Predictive Marketing

PredictiveMarketing

PredictiveMarketingEasy Ways Every Marketer

Can Use Customer Analyticsand Big Data

Ömer Artun, PhDDominique Levin

This book is printed on acid-free paper. ♾

Copyright © 2015 by AgilOne. All rights reserved

Published by John Wiley & Sons, Inc., Hoboken, New JerseyPublished simultaneously in Canada

No part of this publication may be reproduced, stored in a retrieval system, or transmitted in any formor by any means, electronic, mechanical, photocopying, recording, scanning, or otherwise, except aspermitted under Section 107 or 108 of the 1976 United States Copyright Act, without either theprior written permission of the Publisher, or authorization through payment of the appropriateper-copy fee to the Copyright Clearance Center, 222 Rosewood Drive, Danvers, MA 01923,(978) 750-8400, fax (978) 646-8600, or on the web at www.copyright.com. Requests to the Publisherfor permission should be addressed to the Permissions Department, John Wiley & Sons, Inc.,111 River Street, Hoboken, NJ 07030, (201) 748-6011, fax (201) 748-6008, or online atwww.wiley.com/go/permissions.

Limit of Liability/Disclaimer of Warranty: While the publisher and author have used their best effortsin preparing this book, they make no representations or warranties with the respect to the accuracy orcompleteness of the contents of this book and specifically disclaim any implied warranties ofmerchantability or fitness for a particular purpose. No warranty may be created or extended by salesrepresentatives or written sales materials. The advice and strategies contained herein may not besuitable for your situation. You should consult with a professional where appropriate. Neither thepublisher nor the author shall be liable for damages arising herefrom.

For general information about our other products and services, please contact our Customer CareDepartment within the United States at (800) 762-2974, outside the United States at (317) 572-3993or fax (317) 572-4002.

Wiley publishes in a variety of print and electronic formats and by print-on-demand. Some materialincluded with standard print versions of this book may not be included in e-books or inprint-on-demand. If this book refers to media such as a CD or DVD that is not included in theversion you purchased, you may download this material at http://booksupport.wiley.com. For moreinformation about Wiley products, visit www.wiley.com.

Library of Congress Cataloging-in-Publication Data:

Artun, Omer, 1969–Predictive marketing : easy ways every marketer can use customer analytics and big data /

Omer Artun, Dominique Levin.pages cm

Includes index.ISBN 978-1-119-03736-1 (hardback)ISBN 978-1-119-03732-3 (ePDF)ISBN 978-1-119-03733-0 (ePub)

1. Marketing. I. Levin, Dominique, 1971– II. Title.HF5415.A7458 2015658.8—dc23

2015013473

Cover image: WileyCover design: Abstract Shoppers © Maciej Noskwoski/GettyImages

Printed in the United States of America

10 9 8 7 6 5 4 3 2 1

http://www.copyright.com

http://www.wiley.com/go/permissions

http://booksupport.wiley.com

http://www.wiley.com

Dedicated to

My darling wife Dr. Burcak Artun for always believing in me

Ömer Artun

My husband Eilam Levin without whom it would not be worthwhile

Dominique Levin

CONTENTS

Introduction: Who Should Read This Book ix

PART 1 A Complete Predictive Marketing Primer 1

Chapter 1 Big Data and Predictive Analytics Are NowEasily Accessible to All Marketers 3

Chapter 2 An Easy Primer to Predictive Analyticsfor Marketers 23

Chapter 3 Get to Know Your Customers First: BuildComplete Customer Profiles 43

Chapter 4 Managing Your Customers as a Portfolioto Improve Your Valuation 63

PART 2 Nine Easy Plays to Get Started withPredictive Marketing 75

Chapter 5 Play One: Optimize Your Marketing SpendingUsing Customer Data 77

Chapter 6 Play Two: Predict Customer Personas andMake Marketing Relevant Again 93

Chapter 7 Play Three: Predict the Customer Journeyfor Life Cycle Marketing 103

Chapter 8 Play Four: Predict Customer Value andValue-Based Marketing 115

Chapter 9 Play Five: Predict Likelihood to Buy or Engageto Rank Customers 123

vii

viii Contents

Chapter 10 Play Six: Predict Individual Recommendationsfor Each Customer 137

Chapter 11 Play Seven: Launch Predictive Programsto Convert More Customers 145

Chapter 12 Play Eight: Launch Predictive Programsto Grow Customer Value 155

Chapter 13 Play Nine: Launch Predictive Programsto Retain More Customers 169

PART 3 How to Become a True PredictiveMarketing Ninja 183

Chapter 14 An Easy-to-Use Checklist of PredictiveMarketing Capabilities 185

Chapter 15 An Overview of Predictive (and Related)Marketing Technology 197

Chapter 16 Career Advice for Aspiring Predictive Marketers 209

Chapter 17 Privacy and the Difference Between Delightfuland Invasive 215

Chapter 18 The Future of Predictive Marketing 221

Appendix: Overview of Customer Data Types 229

Index 237

INTRODUCTION: WHO SHOULD READTHIS BOOK

This book is for everyday marketers who want to learn what predictivemarketing is all about, as well as for those marketers who are ready

to use predictive marketing in their organizations. Whether you are justgetting started with your research, or have already begun to implementpredictive marketing, you will find many practical tips in this book.

We share what marketers at companies large and small should knowabout predictive marketing. We show you how to achieve the same largereturns as early adopters such as Harrah’s Entertainment, Amazon, andNetflix. We also give you a practical guidebook to help you get startedwith this new way of marketing. And above all, we share stories fromcompanies small and large, from retail to publishing, to software to man-ufacturing. All of these marketers have achieved revolutionary returns,and so can you.

About This Book

We are passionate about improving the quality of marketing and aboutarming marketers with the knowledge and tools they need to make mar-keting relevant again. We hope that the chapters that follow give mar-keters the vocabulary and the inspiration to start to understand and usebig data and machine learning–powered marketing. We believe this willlead to a win-win for customers, businesses, and marketers. Customerswill have more relevant and meaningful experiences, businesses will beable to build more profitable customer relationships, and marketers willgain visibility and respect within their organizations. We look forward tocontinuing the dialogue on our website www.predictivemarketingbook.com, the “Predictive Marketing Book” LinkedIn group (https://www.linkedin.com/groups?gid=8292127), or via twitter.com/agilone.

ix

http://www.predictivemarketingbook.com

https://www.linkedin.com/groups?gid=8292127

https://www.linkedin.com/groups?gid=8292127

x Introduction: Who Should Read This Book

This book is divided in three main parts. The first part, “A CompletePredictive Marketing Primer,” introduces many of the foundational ele-ments in predictive marketing, including what is happening under thehood of predictive marketing software, how data science and predictiveanalytics work, and what are fundamentals behind the customer life-time value concept. The second part of the book, “Nine Easy Plays toGet Started with Predictive Marketing,” is a playbook with concretestrategies to get you started with predictive marketing. The last part ofthe book, “How to Become a True Predictive Marketing Ninja,” givesan overview of predictive marketing technologies, some career advicefor marketers, and looks at privacy and the future of predictive mar-keting. Many of the chapters can be read as stand-alone essays, so usethe executive summary below to jump to the chapters that are mostrelevant to you.

What Is in This BookChapter 1: Big Data and Predictive AnalyticsAre Now Easily Accessible to All Marketers

Predictive marketing is a new way of thinking about customer relation-ships, powered by new technologies in big data and machine learning,which we collectively call predictive analytics. Marketers better pay atten-tion to predictive analytics. Applying predictive analytics is the biggestgame-changing opportunity since the Internet went mainstream almost20 years ago. Although some large brands have been using pieces of pre-dictive marketing for many years now, we are still in the early stagesof adoption, and this is the right time to get started. The adoption ofpredictive marketing is accelerating among companies large and smallbecause: (a) customers are demanding more meaningful relationshipswith brands, (b) early adopters show that predictive marketing deliversenormous value, and (c) new technologies are available to make predictivemarketing easy.

Chapter 2: An Easy Primer to PredictiveAnalytics for Marketers

Many marketers want to at least understand what is happening in the pre-dictive analytics black box, to more confidently apply these models or to

Introduction: Who Should Read This Book xi

be able to communicate with data scientists. After reading this chaptermarketers will have a good understanding of the entire predictive analyt-ics process. There are three types of predictive analytics models that mar-keters should know about: unsupervised learning, supervised learning,and reinforcement learning. Many marketers don’t realize that 80 percentof the work associated with predicting future customer behavior is goingtowards collecting and cleaning customer data. This data janitor work isnot glamorous but essential: without accurate and complete customerdata, there can be no meaningful customer analytics.

Chapter 3: Get to Know Your Customers First:Build Complete Customer Profiles

Building complete and accurate customer profiles is no easy task, but ithas a lot of value. If yours is like most companies, customer data is all overthe place, full of errors and duplicates and not accessible to everyday mar-keters. Fortunately, predictive technology, including fuzzy matching, canhelp—at least some—to clean up your data mess and to connect onlineand offline data to resolve customer identities across the digital and physi-cal divide. Just getting all customer data in one place has enormous value,and making customer profiles accessible to customer-facing personnelthroughout the organization is a great first step to start to deliver betterexperiences to each and every customer.

Chapter 4: Managing Your Customers asa Portfolio to Improve Your Valuation

It is our strong belief that the best way for any business to optimizeenterprise value is to optimize the customer lifetime value of each andevery customer. Customers are the unit of value for any company andtherefore customer lifetime value is the most important metric in mar-keting. If you maximize the lifetime value, or profitability, of each andevery customer, you also maximize the profitability and valuation of yourcompany as a whole. The best way to optimize lifetime value for all cus-tomers is to manage your customers as if they were a stock portfolio.You take different actions and send different messages for customerswho are brand-new than for those who have been doing business withyou for a while. You will need to adjust your thinking and budget forunprofitable, medium-value, and high-value customers.

xii Introduction: Who Should Read This Book

Chapter 5: Play One: Optimize Your MarketingSpending Using Customer Data

When asked to allocate marketing budgets, most marketers immediatelythink about acquisition spending and about allocating budget to thebest performing channels and products. However, the predictive mar-keting way to allocate spending is based on allocating dollars to the rightpeople, rather than to the right products or channels. Most companiesare focused on acquisition, whereas they could achieve growth morecost-effectively by focusing more of their time and budget on retentionand reactivation of customers. Marketers should learn to allocate budgetsbased on their goals to acquire, retain, and reactivate customers and tofind products and channels that deliver the highest value customers.

Chapter 6: Play Two: Predict Customer Personasand Make Marketing Relevant Again

We will look at the predictive technique of clustering and how it isdifferent from classical customer segmentation. Clustering is a power-ful tool in order to discover personas or communities in your customerbase. Specifically, in this chapter we look at product-based, brand-based,and behavior-based clusters as examples. Clustering can be used to gaininsight into differences in customers’ needs, behaviors, demographics,attitudes, and preferences regarding marketing interactions, products,and service usage. Using these clusters, you can also start to differentiateand optimize both marketing actions and product strategy for differentgroups of customers.

Chapter 7: Play Three: Predict the CustomerJourney for Life Cycle Marketing

In this chapter we look at the customer life cycle in more detail, fromacquisition, to growth, and to retention and see how your engagementstrategy should evolve with each and every customer during the lifecycle. The basic principle of optimizing customer lifetime value is thesame for all stages of the life cycle and can be summarized in three words:give to get. Customers are much more likely to buy from you if they trustyou. The best way to gain trust is to deliver an experience of value. So toget customer value, give customer value.

Introduction: Who Should Read This Book xiii

Chapter 8: Play Four: Predict Customer Valueand Value-Based Marketing

Not all customers have equal lifetime value. Any business will havehigh-value customers, medium-value customers, and low lifetime valuecustomers. There is an opportunity to create enterprise value by craftingmarketing strategies that are differentiated based on the value of the cus-tomer. This practice to segment and target by customer lifetime value iscalled value-based marketing. Spend more money to appreciate and retainhigh-value customers. Upsell to medium-value customers in order tomigrate these customers to higher value segments. Finally, reduce yourcosts to service low-value or unprofitable customers.

Chapter 9: Play Five: Predict Likelihood to Buyor Engage to Rank Customers

Likelihood to buy models is what most people think about when youuse the word predictive analytics. With these models you can predict thelikelihood of a certain type of future behavior of a customer. In thischapter we look at programs based on likelihood to buy predictionsspanning both consumer and business marketing. We see how in busi-ness marketing predictive lead scoring or customer scoring can optimizethe time of your sales and customer success teams. We also show youhow consumer marketers can optimize their discount strategy and thefrequency of their emails based on propensity models.

Chapter 10: Play Six: Predict IndividualRecommendations for Each Customer

Another popular predictive technique is personalized recommendations.In this chapter we provide marketers a primer on recommendations andwe teach you about different types of recommendations. We explorerecommendations made at the time of purchase versus those made as afollow-up to a purchase, and recommendations that are tied to specificproducts versus those that are tied to specific customer profiles. We alsodiscuss what can go wrong when making personalized recommenda-tions, and we highlight the need for merchandising rules, omni-channelorchestration, and giving customers control when making personalrecommendations.

xiv Introduction: Who Should Read This Book

Chapter 11: Play Seven: Launch Predictive Programsto Convert More Customers

In this chapter we cover three specific predictive marketing strategiesthat can help you acquire more, and better, customers: using personas todesign better acquisition campaigns, using remarketing to increase con-version and using look alike targeting. When it comes to remarketing,you should be able to differentiate between customers who are likelyto come back, and send them a simple reminder, versus those who areunlikely to come back and may need an additional incentive. This istrue for abandoned cart, browse, and search campaigns. Using looka-like targeting features of Facebook and other advertising platforms, youcan find more customers who look just like your existing customers, forexample, new customers just like your best customers.

Chapter 12: Play Eight: Launch PredictivePrograms to Grow Customer Value

The secret to retaining a customer is to start trying to keep the customerthe day you acquire her. The initial transaction is just the beginningof a long relationship that needs to be nurtured and developed.Engagement with customers should not stop when you convert aprospect into a buyer. In this chapter we cover a number of specific pre-dictive marketing strategies to help grow customer value: postpurchasecampaigns, replenishment campaigns, repeat purchase programs, newproduct introductions, and customer appreciation campaigns. We willalso discuss loyalty programs and omni-channel marketing in the age ofpredictive analytics.

Chapter 13: Play Nine: Launch PredictivePrograms to Retain More Customers

We recommend you focus on dollar value retention. If you don’t, youcould be retaining customers, but losing money anyway. Also, whenmeasuring customer retention it is important to realize that not all churnis created equal. Losing an unprofitable customer is not nearly as badas losing one of your best customers. Also, it is a lot easier, cheaper,and more effective to try and prevent a customer from leaving than

Introduction: Who Should Read This Book xv

it is to reactivate that customer after she has already stopped shoppingwith you. In this chapter we look at different churn management pro-grams, from untargeted, applying equally to all your customers, to tar-geted, and we will cover proactive retention management and customerreactivation campaigns.

Chapter 14: An Easy-to-Use Checklist of PredictiveMarketing Capabilities

In order to use the predictive marketing techniques discussed in thisbook you need to acquire both a predictive marketing mind-set aswell as certain predictive marketing technical capabilities. You need toevolve your thinking from being focused on campaigns, channels, andone-size-fits-all marketing to being focused on individual customersand their context. From a technology point of view you need to acquirebasic capabilities in the areas of customer data integration, predictiveintelligence, and campaign automation.

Chapter 15: An Overview of Predictive (and Related)Marketing Technology

We live in an exciting and somewhat confusing time. A large number ofnew marketing technologies are becoming available every year. In thischapter, we will give you a high-level overview of the various types ofcommercially available technologies and describe what it would take tobuild a predictive marketing solution in-house from the ground up.

Chapter 16: Career Advice for AspiringPredictive Marketers

There is a huge career opportunity that comes from being an earlyadopter of new methodologies and technologies, predictive marketingand predictive analytics included. If you are uncomfortable with num-bers and math, and fearful of getting started with predictive marketing,there are a couple of things you should know: business understandingtrumps math, asking the right questions goes a long way, the best mar-keters blend the art and science of marketing, and there is a lot you canlearn from others.

xvi Introduction: Who Should Read This Book

Chapter 17: Privacy and the Difference BetweenDelightful and Invasive

In general, consumers are willing to share preference information inexchange for apparent benefits, such as convenience, from using per-sonalized products and services. When it comes to personalization,there are different types of customer information that can be used andconsumers may feel different about one type of information over theother. Use common sense when considering whether a marketing cam-paign is delightful or creepy and consider the context of the situation.This chapter will provide some guidelines for dealing with customer datathat will engender trust.

Chapter 18: The Future of Predictive Marketing

Predictive analytics will continue to find new applications inside andbeyond marketing. Not only will more algorithms become available,but real-time customer insights will start to shape our physical world,including the store of the future. There are huge benefits for customers,companies, and marketers alike to get started with predictive marketingsooner rather than later. Sooner or later your customers and competitorswill force you to adopt a predictive marketing mind-set, so you might aswell be an early adopter and derive a huge competitive advantage.

About the Authors

Omer Artun

I am a scientist by training; I am an entrepreneur at heart, driven bycuriosity of knowledge and challenging status quo. In elementary school,I saw the opportunity to make a profit collecting fruit from mulberrytrees from our school backyard and selling it on the street, enlistingmy schoolmates to help me run this small business. With some prod-ding from my engineer parents, I followed in my older brother’s foot-steps to enter a PhD program in physics at Brown University, studyingunder Leon Cooper at The Institute for Brain and Neural Systems.Dr. Cooper has received the Nobel Prize in Physics for his work onsuperconductivity and later decided that the next big problem to solve

Introduction: Who Should Read This Book xvii

was in neuroscience, decoding how we learn and adapt. He is a pioneerin learning theory since the early 70s, using both experimental neuro-science as a base as well as statistical techniques for understanding andcreating learning systems, now popularly called machine learning. I workedon both biological mechanisms that underlie learning and memory stor-age as well as construction of artificial neural networks, networks thatcan learn, associate, and reproduce such higher level cognitive acts asabstraction, computation, and language acquisition. Although these tasksare carried out easily by humans, they have not been easy to embody asconventional computer program.

As I was getting close to graduating from the PhD program at BrownUniversity around 1998, I noticed that the business world was mostlyrunning on simple spreadsheets, and I wanted to apply a data scienceand machine-learning approach to business. This goal led me to work forMcKinsey & Co., the premier strategy consulting firm that helps largecompanies formulate strategies based on a fact-based problem solvingapproach.

When I joined McKinsey & Co. in 1999, I was able to test drivesome of this data scientific approach in a few studies. My first projectwas to help a large technology company improve sales coverage, scien-tifically matching the sales team with the customers based on customerneeds, sales team’s skill, and experience. The CEO was impressed withthe results on paper, but was unable to operationalize the results in reallife, in a repeatable way. This is what I call the last mile problem ofanalytics. I realized that this is a big problem to solve. Analytics is animportant enabler in improving commercial efficiency, but can only cre-ate value if it becomes part of the day-to-day execution workflow. I sawthis theme repeat over and over again in many areas of business, pricing,supply chain, marketing, and sales. Most McKinsey projects I have beenpart of ended up on a slide deck which had all the right answers butvery rarely created any real value. Equipped with McKinsey training, Ijoined one of my clients, Micro Warehouse as VP of Marketing, in 2002,with the goal to bring data science to everyday operations. I was luckyto be empowered by the CEO Jerry York and President Kirby Myers.Jerry was the most analytically driven person I ever knew in business,still to this day. He was previously CFO of IBM during Gerstner years,and CFO of Chrysler before that. He encouraged me to use data scienceto help him run the business better.

xviii Introduction: Who Should Read This Book

I knew I had to architect my approach in a way that married datascience with execution to solve the last mile problem. I had two impor-tant recruits, Dr. Michel Nahon, a brilliant Yale-trained applied math-ematician who helped me with machine-learning algorithms, and thehacker extraordinaire Glen Demeraski, who helped me with everythingdatabase and application related. I created approaches and systems thatused data to more efficiently allocate resources, reduce marketing costs,and uncover new revenue sources. We had significant impact on mar-keting efficiency, pricing, and discounting patterns as well as salesforceeffectiveness. In early 2003 we had real-time systems alerting purchase,pricing, and customer acquisition patterns of the sales team compared tomoving averages to take immediate action by the sales leadership. AfterMicro Warehouse, from 2004 to 2006, I joined Best Buy as Senior Direc-tor of Business-to-Business marketing of its newly founded Best Buy forBusiness division. Best Buy at the time also struggled with the same exactlast mile problem, lots of internal resources, tools, many high-flying con-sultants talking about customer segmentation, and analytics, but whenyou walked into a store, none of that had any impact at the customerlevel. This is the true test of analytics; does it impact the customers in apositive way that they can experience it? If not, then you have the wrongsetup. Making progress at Best Buy was much more difficult, which I willtouch on in Chapter 1.

While working at Micro Warehouse and Best Buy, I was also aregular guest lecturer at Columbia University and NYU Stern MBAprograms Relationship Marketing and Pricing courses that Dr. HitendraWadhwa taught. I also became an Adjunct Professor at NYU Stern forSpring 2006, teaching the MBA level Relationship Marketing program.During this period, talking to students, doing market research, talking tocolleagues at different companies, I postulated that data-driven predic-tive marketing would become the new paradigm for the next 10 years.The value of predictive marketing was already clear to me, but its impor-tance has accelerated due to digital transformation of commerce, increasein customer touch-points, and exponential increase in the size, variety,and velocity of data (which is now popularly called “big data”).

If you ask me what is the one important thing I learned fromDr. Cooper, I would say that it is breaking the problem down to its coreand solving it at a fundamental level. He always said the idea behind thesolution to any problem has to be clean and very simple. This is how I

Introduction: Who Should Read This Book xix

thought about the marketer’s problem. Marketing was easy in the daysof the old corner store. People knew our name, our likes and dislikes,and treated us on a one-to-one basis. Marketers lost touch with theircustomers in the era of one-size-fits-all mass optimization. Customersbecame survey responders and focus group participants; it was allabout products and channels. However, the need for customer-centricmarketing has always been there, it just wasn’t practical and cost effectiveto practice. Digital transformation including web, email, mobile, social,location technologies combined with technologies to store, process,and extract information has significantly changed what is practical andcost effective.

Predictive marketing is the approach that restores that personal touchby bringing that human sensibility into our digital and offline lives, byfocusing on the consumers individually to understand what they did andwhat they will do next. Predictive analytics, based on machine-learningalgorithms, offers enormous leverage to marketers trying to make senseof these actions. Rather than replacing human decision making, machinelearning and complex algorithms could help people amplify their intel-ligence and deal with problems on a much larger scale, something likegiving a bulldozer to people used to digging with a shovel.

I saw the opportunity to solve a problem that a growing numberof companies were struggling with, and I decided to disrupt the statusquo and solve this problem. In 2006, I founded AgilOne, to bring thepower of big data and predictive analytics to everyday marketers with aneasy-to-use, yet powerful, cloud-based software platform.

AgilOne was initially bootstrapped for the first 5 years, then backedby top tier VC firms including Sequoia Capital, Mayfield Fund, TenayaCapital, and Next World Capital. We are helping more than 150brands in retail, B2B, Internet, media, publishing, and education deliverrelevant experiences across channels. Through complete and accuratecustomer profiles, predictive insights, and built-in life cycle marketingcampaigns, marketers boost customer loyalty and increase customerlifetime value.

In my spare time, I claim to be an accomplished potter of 28 years,having studied at Rhode Island School of Design under Lawrence Bushduring my years at Brown. A native of Turkey, I now live in Los Gatoswith my wife Burcak and two daughters, Ayse and Leyla. As I write thisintroduction, my daughter Ayse, who is a freshman at Castilleja School

xx Introduction: Who Should Read This Book

in Palo Alto, is reading an article about predictive marketing for her mathclass, which shows how predictive marketing will become mainstreamfor the next generation.

Dominique Levin

I credit my education, a combination of engineering school, designschool, and business school for my left-brain–right-brain approach tomarketing: I have a master’s of science (Cum Laude) in industrial designengineering from Delft University in The Netherlands and a master’sof business administration (with Distinction) from Harvard University.I recommend all marketers to marry human creativity with technologylearning in order to deliver value to customers. Over the past 20 yearsI have run marketing at companies large and small, on four differentcontinents, targeting businesses and consumers. Above all, I was an earlyconvert to the importance of customer data.

In 1994 I took my first marketing job: a summer internship inCusco, Peru. I drove around in a pickup truck to visit local farmers andtally how many would join a local cooperative to process fruits into mar-malades and liquors. For my next job, at Philips Consumer Electronics,I was asked to find a way to sell more electronics to girls and women.I mingled with teenagers at local high schools to collect data. Philipslaunched a product called KidCom, an electronic organizer for girls, andproto-typed TeenCom, a two-way paging device for teenagers. My bosson this project was Tony Fadell, who later became the father of the iPodand iPhone, and who went on to found NEST. In 1997, I relocatedto Tokyo, Japan, to work for Nippon Telegraph and Telephone(NTT). All employees at NTT, whether in product or finance, workedone weekend in the company store to meet and serve customers.I recommend such “meet the customer” program to any company asno numbers can totally replace meeting customers face to face.

In 2000, I moved to Silicon Valley and ran marketing for my first bigdata company, LogLogic—later acquired by TIBCO Software. For thefirst time I had access to lots of customer data in digital form. Log files arelike the digital video cameras of the Internet. At LogLogic we used thislog data to monitor security, but it also opened my eyes to the possibilitiesof using similar data to better understand and serve customers.

Introduction: Who Should Read This Book xxi

I went on to work for several other technology companies, includingFundly and Totango, focusing on building highly data-driven market-ing organizations. Fundly helps non-profits use social media to raisemoney. We used data to automate the process from self-service sign-up tofundraising success. Totango offered a predictive marketing solution thatmonitors customer behavior to identify both promising and strugglingcustomers. In both cases data and predictions helped to accelerate cus-tomer acquisition and increase customer lifetime value, while loweringthe cost of sales.

I met Omer in my role as CMO at Agilone, where I got to workwith thousands of marketers just like you to figure out how they canbest use customer data to delight customers. Omer and I are united inour data-driven and customer-centric approach to marketing. Data andhumanistic experiences go hand-in-hand. Our passion for customers hasled us to this book.

In my spare time, I love to travel with my husband and three childrenand experience people, places, and cultures around the world. I playice hockey to blow off steam and was once a member of the Dutchnational team. I love to work with entrepreneurs and help them maketheir dreams a reality.

Acknowledgments

This book was significantly enhanced by the efforts of Anne Puyt,Barbara Von Euw, Rinat Shimshi, Dhruv Bhargava, Carrie Koy, JoeMancini, Angela Sanfilippo, Hac Phan, and Francis Brero, who notonly work tirelessly every day to help companies be successful withpredictive marketing, but who also went above and beyond the call ofduty to add their experiences, examples, and wisdom to the manuscript.

We also want to thank visionary CEOs and CMOs who wereearly adopters of the predictive marketing approach, specifically JohnSeabreeze, VP Marketing at Billy Casper Golf; Joe McDonald, SVPSales and Marketing of Stargas, Eoin Comerford, CEO of Moosejaw;Levent Cakiroglu, CEO of Arcelik; Ersin Akarlilar, CEO of Mavi;Adam Shaffer, EVP Marketing of TigerDirect.

Additionally, Omer’s personal success, the success of AgilOne, andthe concepts in this book would not have become a reality without the

xxii Introduction: Who Should Read This Book

help from Bonnie Bartoli, Peter Godfrey, and his “adopted sons anddaughter” Ozer Unat, Dhruv Bhargava, Oyku Akca, Anselme LeVan,Louis Lecat, Ryan Willette, and Francis Brero.

We would also like to thank our families:Omer would also very much like to thank his wife Dr. Burcak Artun,

always believing and encouraging him for challenging the status quo andbeing patient with his busy schedule.

Dominique thanks her husband, Eilam, and children Liv, Yanai, andMilo, for their encouragement during the writing process. Similarly,she would like to thank her AgilOne marketing superstars, Chris Field,Johnson Kang, Kessawan Lelanaphaparn, and Angela Sanfilippo for beingso independent and professional so she could focus on the book at times.

PART 1

A CompletePredictiveMarketing Primer

1

CHAPTER 1

Big Data and PredictiveAnalytics Are NowEasily Accessible to AllMarketers

Predictive marketing is the evolution of relationship marketing definedand practiced by many direct marketers in the last few decades.

Predictive marketing is not a technology, but an approach or a philos-ophy. Predictive marketing uses predictive analytics as a way to delivermore relevant and meaningful customer experiences, at all customertouch points, throughout the customer life cycle, boosting customerloyalty and revenues.

The rise of predictive marketing is fueled by three factors:(1) customers are demanding a more personal, integrated approach asthey interact with marketing and sales through many channels, (2) earlyadopters show that predictive marketing delivers enormous value, and(3) new technologies are available to capture new and existing sources ofcustomer data, to recognize patterns, and to make it easier than ever touse customer data at the intersection of the physical and digital worlds.

Predictive analytics is a set of tools and algorithms used to makepredictive marketing possible. It is an umbrella term that covers a vari-ety of mathematical and statistical techniques to recognize patterns in

3

4 A Complete Predictive Marketing Primer

data or make predictions about the future. When applied to marketing,predictive analytics can predict future customer behavior, classify cus-tomers into clusters among other use cases. Other terms you might hearin the media to describe this process include machine learning, patternrecognition, artificial intelligence, and data mining. Predictive analytics andmachine learning are used interchangeably in this book.

Predictive marketing is fundamentally changing both business andconsumer marketing across the customer life cycle. It is transforming thefocus from products and channels to a focus on the customer. Predictiveanalytics is used to improve strategies to acquire new customers, to growcustomer lifetime value, and to retain more customers over time.

Innovative, technology driven companies like Netflix and Amazonhave been using predictive analytics for years, and so have others likemany in the telecommunications, financial services, and gaming indus-tries, such as Harrah’s Entertainment. The row of movies and TV shows“you might like” that appear when you curl up on the couch and turn onNetflix is a driving force of the company’s success. It’s all made possibleby the translation of customer data with smart analytics. In fact, “75% ofwhat people watch [on Netflix] is from some sort of recommendation,”Netflix’s Research Director Xavier Amatriain wrote on the company’stech blog in 2012.

Amazon has been using predictive analytics to drive success since thevery beginning of the company. Recommendations that appear undera product you are thinking of adding into your cart is part of whatmakes Amazon such an e-commerce powerhouse today. The companyhas stated publicly that 35 percent of its sales comes from recommenda-tions made by their predictive engines. That would equate to $26 billionof revenue in 2013. The company is using predictive analytics in manyother ways too, such as predicting which email newsletter to send you,or to nudge you at the right times to reorder an item.

In the gaming industry predictive models can set budgets andcalendars for the casino’s gamblers, calculating their predicted lifetimevalue in the process. If a gambler wagers less than usual because they mayhave skipped a monthly visit, the casino can intervene with a letter orphone call offering a free meal, a show ticket, or gaming comps. Withoutthis type of customer analytics, casino operators might not noticewhat could be a slight, almost imperceptible change in customer behav-ior that might portend future problems with that patron. For example,

Big Data and Predictive Analytics Are Now Easily Accessible to All Marketers 5

if a long-time customer decides to cash in all their player card points,perhaps it’s because they are dissatisfied with their last experience atthe casino property. Predictive analytics can quickly spot these trendsand alert casino management to the issue so that they can approach theindividual to find out if there is a problem. This kind of personalizationcan go a long way in appeasing a disgruntled customer, which might bethe difference between retaining or losing them as a customer.

Harrah’s Entertainment’s Total Rewards, which was rolled out asTotal Gold in 1997 and renamed Total Rewards a year later, is heraldedby many as the gold standard of customer-relationship programs andis powered heavily by predictive analytics algorithms. The company’sbelief in its loyalty program grew so strong that it cut its traditional adspending from 2008 and 2009 more than 50%. The company spent $106million on measured media in 2008; for the first half of last year it spent$52 million and in this year’s first half $20 million. (Source: http://adage.com/article/news/harrah-s-loyalty-program-industry-s-gold-standard/139424/.)

Although some large brands have been using predictive analyticsfor many years now, it is not too late for other brands, large and small.In fact, predictive marketing is only now finding widespread adoptionin medium and small organizations. A good example of a companythat has achieved significant success with predictive marketing is Mavi,a high-fashion clothing manufacturer and retailer based in Istanbul,Turkey. Mavi is known for its organic denim favored by celebrities andsupermodels. Mavi operates over 350 multinational stores and sales chan-nels in the United States, Canada, Australia, Turkey, and 10 Europeancountries.

Mavi started with a single predictive marketing campaign six yearsago. When Mavi first got started, each department, including mar-keting and IT, used its own set of marketing reports and customerdata, including key performance indicators. This led to cumbersomecross-referencing and impeded important decision making. Like manycompanies, the Mavi marketing team initially did not have access to cus-tomer data without relying on IT resources. This was the first problemthat the team tackled. Mavi deployed a modern, cloud-based predictivemarketing solution in 2009. This allowed the company to consolidate,cleanse, and de-dupe their customer data on a daily basis. They werethen ready to start using data in hyperpersonalized campaigns.

http://adage.com/article/news/harrah-s-loyalty-program-industry-s-gold-standard/139424/


One of the first predictive marketing programs that Mavi tested was aprogram around specific buying personas. Mavi used predictive analyticsto find groups of people with distinct product preferences. In predictivelingo these are called product-based clusters. Mavi found at least three verydifferent groups of shoppers: customers who favored mostly wovenshirts, others who favored beachwear, whereas a third persona mostlyshopped for new season high fashion and accessories. Mavi started touse these personas to implement more targeted marketing campaignsvia email and short message service (SMS). Specifically, it implementeda reengagement campaign for lapsed customers that featured theright types of products with the right customers. Using these clusters,Mavi was able to reactivate 20 percent of lapsed customers. This was a bigbreakthrough because every customer saved or reactivated reduces Mavi’sneed to acquire new customers.

Mavi today is running more than 80 different predictive marketingprograms in a year. Collectively, these campaigns helped add 7 percentagepoints to Mavi’s overall revenues in the first few years, which is a hugesum on a dollars and cents basis. Wikipedia reports that Mavi revenuesin 2014 were $747 million, so that would be an incremental $52 million.Mavi is still finding new ways to increase customer lifetime value, andwith every campaign launched this number is pushed up higher.

Elif Oner, Mavi’s head of customer relationship management, rec-ommends all marketers get started with predictive marketing. She says:“Start small and pick just one program and build on that success.” Elif isalso the CFO’s favorite marketer. Every dollar she spends in market-ing, every discount she gives, is accounted for, tested, and optimized.The CIO Bulent Dursun also played an important role in realizing thepotential of analytics and was a key supporter, which made the approachsuccessful.

The Predictive Marketing Revolution

Anticipating customer needs is not a new concept. What is new is theability to anticipate and respond to customer needs automatically, nearreal time and at large scale, for hundreds, thousands, or even millions ofcustomers at a time.

Not too long ago, you could walk into a corner store and the salesper-son would know your name, know what kind of things you bought, how


long you’ve been a customer, and other important information aboutyour personality and behavior. This relationship not only makes the buy-ing process pleasant, it also increases the likelihood of the customer toreturn, spend more, and develop a sense of brand loyalty and trust.

These days we shop in supermarkets where nobody knows our name.The promise of predictive marketing is to bring the personal relationshipsof the corner store to the modern world of online and offline marketing.Using predictive analytics, it is possible to move from an era of massmarketing centered on the products you sell and the promotions yousend to an era of highly personalized marketing centered on the customeryou serve.

Today, even small- and midsize businesses interact with customers onan enormous scale through a wide variety of channels, including web-sites, social media, mobile apps, and store visits. Because of the substantialincrease in speed, number, and type of customer interactions, companieshave a greater opportunity to maintain the kind of personal relationshipsthat used to be an important aspect of doing business. Of course this isnot easy, and many companies fail due to lack of technical, organizationalcapabilities and strategic focus.

Customer interactions and the digitization of so much of our dailyactivities have allowed businesses to gather an extraordinary amountof data about their customers that can be put to use to better servicecustomers. For example, when you buy a pair of shoes at Zappos, thecompany knows many things about you: what type of shoes you like,your name, gender, where you live, whether your zip code is mainlymade up of apartment buildings or single-family homes, whether youtypically buy items at full price or at a discount, whether you bought justone product or multiple products, how often you clicked on a Zapposemail or visited its website before placing that first order and what youlooked at, how often you called the call center, whether you are afirst-time or repeat customer, whether you are a VIP customer or anunprofitable customer who returns more products than she keeps, andmuch more.

Most companies still find it very difficult to put any of this informa-tion to good use. The sheer size and breadth of the records make themincomprehensible for anyone without the training and experience tomine insights from large datasets. This is where predictive analytics andmachine learning come in. Machines are very good at mining insights


Figure 1.1 The Predictive Marketing Revolution

from large datasets automatically. Machines can remember the names ofmillions of customers with no effort and greet them accordingly, justas the shopkeeper from yesteryear would have done. In other words,using machines, humans can now bring back the personalized market-ing interactions from yesteryear, even if their company has millions ofcustomers. Figure 1.1 illustrates how the marketing revolution has comefull circle. In the 1800s, shopkeepers had personal relationships witheach and every customer. In the 1900s, during the industrial revolution,these personal relationships fell victim to mass marketing and a desireto scale businesses. Now, thanks to the technological revolution, mar-keters can bring back the personal relationships from yesteryear, whilestill operating companies at a large scale.

Predictive marketing is the perfect marriage between machine learn-ing and human intelligence. The point of predictive marketing is not toreplace marketers with machines but rather to empower and augmenthuman intelligence with machine learning.

The Power of Customer Equity

Predictive marketing gives rise to a new, data-driven way to approachmarketing, with the customer at its center. The ability to collect andanalyze data on every single customer, as well as his or her interactions


Figure 1.2 From a Product to a Customer Orientation

with your brand, allows you to serve your customers better and generatemore sales. At its core, as Figure 1.2 illustrates, predictive marketing ishelping companies to evolve from a product- or channel-centric orien-tation to a customer-centric orientation. Companies using predictivemarketing focus on developing and managing customer relationshipsrather than just developing and selling products or channels:

• Instead of finding customers who will want your products, it is nowpossible to discover which products your customers will want in thefuture.

• Instead of maximizing sales, companies in the customer era focuson optimizing customer lifetime value and share of wallet to driveprofitability of the enterprise.


• Instead of organizing around channels and product lines, companieswhich practice predictive marketing organize around the customer.

• With the customer at the center, companies are using big data andpredictive analytics to configure processes and organizations to findways to customize interactions.

• Communications become much more targeted and the key metric isrelevance, not reach.

Predictive marketing allows you to identify and realize the long-termvalue of customer relations to keep your best customers coming back andbuying more. Figure 1.3 illustrates the core principle: if your companyacquires more profitable customers, grows the value of each and everycustomer systematically, and retains these customer relationships for along time, the firm will grow, too.

Companies should think about managing customer equity in muchthe same way they manage their stock portfolios: just like stocks,some customers are more valuable than others and their value will riseand fall throughout time. Predictive marketing gives companies an easyand automated way to manage individual customer lifetime value andcustomer equity.

The key to unlocking this value lies in the information you are ableto collect about your customers. The more you can personalize the expe-riences you offer, the more likely the customer remains loyal to yourbrand. Think about your hairdresser. She has a lot of information aboutyou. She knows how you like your hair cut and probably knows a lotabout your family, friends, and job. This information makes the inter-action with your hairdresser very seamless. You sit down, she gets towork, and you have a pleasant conversation. She may call you when it’stime for your next appointment and suggest a new hairstyle from time totime. It would take you a long time to start over with a new hairdresser.Your hairdresser has very few clients. Most marketers serve millions ofcustomers. It is not possible for a brand to collect and process the dataof millions of customers without computers and software.

Predictive marketing puts customer data and insights directly in thehands of marketers, customer-facing personnel, and applications thatdeliver personalized experiences to individual customers.


Figure 1.3 Customers Are Key to Market Value

Predictive Marketing Use Cases

Predictive marketing is much more than just providing recommenda-tions. The most commonly use cases of predictive marketing are thefollowing:

• Improve precision of targeting and acquisition efforts. Withpredictive marketing it is possible to know which channels producethe most profitable customers and optimize marketing spending based


on this knowledge. Armed with better information about behavioralbuying personas, marketers can also design more effective acquisi-tion campaigns that hypertarget a specific microsegment and increaseconversions by four times or more.

• Use personalized experiences to increase lifetime value.Predictive marketing can predict future customer preferences andinteractions (such as a customer’s likelihood to buy). Armed withthis information, marketers can improve personalization, relevancy,and timing of customer interactions. It is these experiences that willkeep customers coming back and maximize customer lifetime value.If you can maximize the lifetime value of each of your customers,you will automatically maximize the value of your entire customerportfolio and thereby the value of your company as a whole.

• Understand customer retention and loyalty. Predicting when,why, and which customers will return or leave is a big challengefor many organizations. Predictive marketing can help flag customerswho are at risk of leaving so that marketers can take proactive stepsto retain these customers. Predictive analytics can also generateinsights about loyalty-inducing behaviors that maximize customerlifetime value.

• Optimize customer engagement. Predicting who will respond toan email promotion, what would it take to convert a browser into abuyer, what discount is needed to incent the customer to completethe transaction are all methods of increasing customer engagement inreal time or near real time that maximizes marketing effectiveness.

Figure 1.4 gives examples of questions that predictive analytics cananswer for marketers. These examples, as well as other used cases, arediscussed in greater detail throughout the book. The list below is not anall-inclusive list, as the marketing questions that can be answered withpredictive analytics are really endless.

Armed with information ranging from likelihood to buy, predictedlifetime value, and future product preferences, brands can betterserve their prospects and their customers by delivering personalizedexperiences.


10 Questions to Answer How Predictive Can Help

1. Who my best customerswill be

Predict which prospects or customers have the highestlifetime value, taking into account revenues, butalso the cost to acquire and service these accounts.Use this information to spend time and money onhigh-potential customers early on.

2. Find more new customerslike your existing bestcustomers

Predict which prospects are most like your existinghigh-value customers using look-alike targeting(B2C) or specialized lead generation vendors(B2B).

3. Find personas in your data touse to acquire morecustomers like this

Predict the customer clusters that most distinguishbuying personas with respect to brands, products,content and behaviors in your customer base.Then develop creative, content, products, andservices to attract more buyers like this.

4. Which marketing channelsare most profitable

Predict which channels attract the customers with thehighest lifetime value, including all futurepurchases. Use this information to influencekeyword bidding strategies and channelinvestments.

5. Which prospects (nonbuyers)are most likely to buy

Determine who is most likely to buy so you can givethe right incentive (in B2C) or prioritize your salespersonnel’s time with the right prospects (in B2B).

6. Which existing (or past)customers are most likelyto buy

Product incentive (or discount) is needed to convincea one-time buyer to become a repeat customer.Prioritize the time of account managers to focuson likely upsell candidates.

7. Which existing customersare least likely to buy

Predict which customers are likely to leave and targetthem proactively with a “please come back”incentive, a personalized recommendation or byhaving the customer success manager make a call.

8. What customers might beinterested in a specific newproduct

Predict which customers might be interested inoverstock items or a new product release so youcan focus your sales and marketing efforts on thesebusinesses or consumers.

9. What other products orcontent might this customerbe interested in

Predict what product or content recommendations tomake to a particular customer in order to win,upsell, or reengage this customer.

10. What is my share of walletwith a specific customer

Predict in what markets or customer groups you havehigh value potential to focus future customeracquisition strategies.

Figure 1.4 Ten Examples of Predictive Marketing


Predictive Marketing Adoption Is Accelerating

A recent survey of 132 marketing executives by our company AgilOnefound that 76 percent of marketers used some form of predictiveanalytics in their marketing in 2015, which is up from 69 percent in2014. The acceleration is fueled by three factors: (1) customers aredemanding the benefits of predictive marketing—mainly highly relevantand timely marketing, (2) early adopters show that predictive marketingdelivers enormous value, and (3) new technologies are available to makepredictive marketing easy.

Customers Are Demanding More MeaningfulRelationships with Brands

Consumers are bombarded with marketing and frankly are fed up.Retail research agency Conlumino conducted a consumer survey in late2014 that showed many consumers have come to expect some form ofpersonalization—in part because the larger and more established brandshave been serving up personalized experiences for some years now.By asking more than 3,000 adult online shoppers about what informa-tion they expected companies to know about them and what personal-ized experiences they appreciate, the survey uncovered that more than70 percent of shoppers want brands to deliver some type of personalizedexperience, whether it is sending an alert about a new product thatmatches their interests, a refill reminder, or VIP customer recognition.Certain types of customized experiences, such as loyalty rewards andpersonalized discounts, were popular across the board, whereas appreci-ation levels for other areas of personalization differed greatly dependingon age, location, gender, and a number of other factors. The findingssuggest that it is crucial to have a deep understanding of your customers,and using hypertargeting is crucial to building brand loyalty:

• More than 79 percent of U.S. consumers and 70 percent of U.K.consumers expect some sort of personalization from brands.

• More than half of consumers in the United States and UnitedKingdom expect e-commerce sites to remember their past purchases.

• Among U.S. shoppers, the most popular personalized experienceswere emails offering discounts on products they previously viewed


(66 percent), alerts when products they like are on sale (57 percent),and VIP customer appreciation rewards (51 percent).

• Consumers in the United States were much more likely to expectonline retailers to personalize experiences than those in the UnitedKingdom: about half of Americans want to receive a new customerwelcome greeting, versus only 34 percent in the United Kingdom.

• Shoppers age 18 to 34, part of the “millennial” generation, were morelikely to appreciate almost all forms personalization: 52 percent ofmillennials expect brands to remember their birthday as compared to21 percent of those aged 65-plus.

• Personalization of email is much more popular than personalization ofdisplay advertising, with 66 percent of U.S. consumers and 57 percentof U.K. consumers welcoming email-retargeting, but only 24 percent(U.S.) and 17 percent (U.K.) welcoming web-based retargeting.

In one case, the customers of a high-fashion brand from New Yorkactually wrote to tell the company they felt they were not receiving thepersonalized experience they deserved. Specifically, this company wasconducting postpurchase surveys after each shipment. Some customerswrote that they were frequent shoppers of the brand, yet felt they didnot receive any special treatment. It is rare for customers to express theirdissatisfaction with one-size-fits-all marketing so directly. It is more likelythat customers are letting you know through their actions. Are you expe-riencing an unusually large number of customer complaints, do you havea small number of repeat buyers, or are you seeing a large number ofopt outs from your email campaigns? All of these could be signs thatcustomers are not getting the personal attention they expect.

Another example comes from a small kitchenware company. Foryears, its products were offered in limited quantities and geographic areas.Word spread about the unique products, and to meet customer demand,the products are now offered through its website directly to consumersand in large retail outlets such as Costco. The passionate customer basewas demanding more relevant communications. Customers did not writeor call to tell the company about this, but rather it started to experiencea rising number of opt outs when sending email. Clearly customers weresaying that the one-size-fits-all email campaigns were not suiting theirneeds. Today customers receive much more relevant and timely email,


such as replenishment reminders to reorder barbecue pellets for grillingjust around the time they were running out of their last order. Predictivemarketing has increased the purchase rate from their marketing emailsfrom 1 percent to 4 percent, while reducing the unsubscribe rate by 40percent within just six weeks.

Many marketers may think they are delivering relevant experiences,but consumer perception is often very different. A 2013 AgilOne surveyof 2,000 marketers and consumers, found that 75 percent of marketersbelieve that they are sending as many as 15 relevant marketing campaignsto consumers each year. However, 34 percent of consumers say theycannot remember a single relevant campaign from the past year. Clearlythere is a disconnect between marketers and consumers. The same surveyfound that 52 percent of marketers send the exact same email to all oftheir customers and 65 percent send the exact same number of emails toeach of their customers regardless of their preferences.

Marketers need to change their thinking dramatically. Today, mar-keters may cheer when one of their email campaigns receives a 4 percentclick-through rate. In reality that means that 96 percent of customersdeemed this email irrelevant. That is a terrible result. We believe allcustomers deserve to be served relevant and respectful communications.Instead of sending 100 messages with a relevancy of 1 percent, marketersshould start sending a single message with a relevancy of 100 percent.

Early Adopters Show That Predictive MarketingDelivers Enormous Value

Marketers better pay attention to predictive analytics. Applying predic-tive analytics is the biggest game-changing opportunity since the Inter-net went mainstream almost 20 years ago, because of the unprecedentedarray of insights into customer needs and behaviors it makes possible.When Bill Gates was asked during a 2013 Sequoia Capital event whatcompany he would start if he were starting out today, he answered withtwo words: machine learning.

In his book, Data Driven Marketing, Kellogg School of Managementfaculty Mark Jeffrey proves that high-performing companies spend sig-nificantly more on data infrastructure than lower performers (16 percentversus 10 percent). High performers were defined as the top 25 percentof the dataset, measured by their excellence at marketing a basket of


financial metrics, which validated the high performers indeed get betterfinancial performance. High performers also spend more on customerequity or retention marketing (14 percent versus 11 percent) and lesson demand generation (48 percent versus 58 percent). He also describesthe success of early adopters of predictive analytics in this book.For example, Earthlink used predictive analytics to identify dissatisfiedcustomers about to churn. Taking proactive steps to contact and retainthese customers helped Earthlink reduce churn by 30 percent. Similarly,the supermarket chain Sainsbury Stores used predictive analytics tocluster its customer base in relevant segments. The company then usedthese segments to remodel its stores and customize the product assort-ment in each store based on these data. As a result, revenues increasedby 12 percent.

There is no limit in the number of campaigns that a companycan develop. Mavi, the mid-market retailer we discussed earlier, hasdeveloped over sixty individual campaigns, using predictive analytics,gradually increasing revenues and profitability.

New Technologies Are Available to MakePredictive Marketing Easy

So why are all marketers’ not using predictive marketing techniquesalready? Until recently the technology needed to collect, analyze, andact on large amounts of customer data for hundreds, thousands, ormillions of customers was inaccessible to most marketers. It was tooexpensive, time-consuming, and cumbersome to invest in the technol-ogy and manpower required to collect and analyze customer data andto deliver customer experiences across channels based on these insights.However, recently predictive analytics has matured to the point that outof the box, standard algorithms and technologies are available, whichmarketers can access without the help of data scientists or softwareengineers. In Chapter 15 we discuss in detail the different off-the-shelftools available to marketers today, which make it significantly cheaper,faster, and easier to use predictive marketing.

Predictive Marketing Is Becoming More Affordable The costsinvolved with predictive marketing can include the money spent onhardware and software technology, as well as the people time spent on


data collection, integration, predictive analytics model developmentand deployment, and the people time needed—from marketers or datascientists—for the ongoing maintenance of these models and the use ofthese models in daily marketing campaigns.

Until recently, the data-warehousing infrastructure alone to collectand store customer data could cost you hundreds of thousands or evenmillions of dollars. In his book about data-driven marketing, MarkJeffrey documents that a small regional retailer with 10 stores, 100,000customers, and 1 terabyte of customer data may need to spend $50,000 to$250,000 to build in-house data warehousing infrastructure. That num-ber rises to $2.5 million for a medium-size retail chain with 400 storesand to $250 million for a large, national retailer with 5,000 stores.Nowadays, cloud-based predictive marketing solutions are available foras little as several thousand dollars a month.

Predictive Marketing Is Becoming Easier to Deploy Whicheversolution you use—an off-the-shelf package or an in-house solution—you will need to collect and integrate customer data into a customerprofile for each customer. In a late 2014 survey of 132 marketing exec-utives, AgilOne found that 68 percent of marketers do not have a singleview of each customer. You probably have a lot of information aboutyour customers already, but this data may reside in many different silos.Most companies have separate databases for online transactions, storetransactions, and phone transactions. Web behavior has its own silo asdoes email behavior, social behavior, and service center interactions.Until recently, it could take months or years to accomplish the dataintegration needed to build a single customer profile and to link anddeduplicate all customer data. Recently, more automated solutions haveemerged that make data integration and data cleansing much easier.These solutions often use standard data models that make it easier andfaster to standardize customer data across channels.

Historically you did not just need in-house infrastructure; you alsoneeded in-house or outsourced data integration specialists and datascientists. Data scientists were needed to build custom models to analyzecustomer data—probably using a predictive analytics workbench tool.These models also needed to be tweaked and tuned periodically tocontinue to deliver accurate results. Data scientists are in short supply,and more than 50 new graduate programs have sprung up across


the United States alone to fill the gap. Fortunately, nowadays manymarketing solutions come with out-of-the-box, built-in models thatare proven and tested by other companies in your industry. Some ofthese models even have self-learning capabilities, which means that theyadjust to your evolving customer data automatically over time withoutrequiring ongoing maintenance from a data scientist.

Predictive Marketing Is Becoming More Accessible to MarketersEven if customer data is available within the organization, it may not beavailable to you, the marketer. This happened to Omer Artun, when hejoined Best Buy as the head of marketing for a new division called “BestBuy for Business” in December 2003:

I joined Best Buy from Micro Warehouse where I had already built anear real-time customer analytics system to track and analyze orderson a daily basis. I was hired to do the same for Best Buy’s new B2Bgroup, which sold products such as routers, printers and computersto small businesses. Best Buy, like many companies, was outsourcingits IT to a third party at the time. If you wanted to talk customerdata with this group, it was $10,000 dollars just to meet. I went tothe IT guys and asked for a raw data dump, nothing else, but gotnothing. After a couple of months, I still wasn’t able to obtain a listof customers who had purchased from us in the past. The data wasavailable somewhere, but not accessible to me. I fought the battle foranother nine months before I finally gave up. I started a company,AgilOne, to make predictive analytics and customer data accessibleto marketers shortly thereafter.

Even having access to data will not drive revenue unless data can beused to deliver more relevant experiences to customers. It can be diffi-cult to integrate and share customer information directly with customer-facing personnel or applications that send out or trigger these types ofcustomer communications. It is not unusual for a company to have alot of customer data, but for marketers to be unable to use this infor-mation to segment customers within their email marketing software, forexample, without complicated, time-consuming, and costly data inte-gration. A new generation of marketing software is becoming available,where predictive insights are accessible as drag-and-drop filters to help


segment and target customers and to include personalized content orrecommendations as dynamic content in emails or advertisements.

What Do You Need for Predictive Marketing?

The fundamental building blocks of a successful predictive marketinginitiative, as summarized in Figure 1.5, are:

1. Continuously learn more and more about your customers: capturedata, build profiles and unify the information on a daily basis.We discuss how to do this in great detail in Chapter 3.

2. Analyze customer information and assess customer preferences andprofitability at a micro (individual/segment) and macro level—bothpast and future. Chapter 2 gives you an overview of the differentpredictive algorithms you can use as a marketer.

3. Leverage customer information to profitably personalize experiencesacross all customer touch points, and to optimize the return on invest-ment of your marketing and sales time and money. Part II of this bookis entirely devoted to these practical applications.

In order to do predictive marketing, you need to develop thesethree capabilities of collecting and integrating customer data, analyzingcustomer data, and delivering relevant customer experiences acrosschannels. You can acquire these capabilities in one of three ways: (1) buildpredictive models yourself using a predictive analytics workbench,

Figure 1.5 The Predictive Marketing Process


(2) outsource customer analytics and predictive marketing campaigns toa marketing service provider, or (3) evaluate and buy a predictive mar-keting solution, such as a predictive marketing cloud or a multichannelcampaign management tool. The first option will cost you millions ofdollars and require you to hire an in-house team. Marketing serviceproviders are typically servicing the Fortune 500, and annual contractsfor full customer data integration, analysis, and personalized campaignprobably start at a quarter million dollars a year. Predictive marketingcloud solutions available at the time that we are writing this book startat about $50,000 a year. Chapter 15 provides more details about eachof these options as well as some criteria to decide which route is bestfor you.

Whatever you do after reading this book, we recommend thatyou get started with predictive marketing in some way or another.Early adopters of predictive marketing will have significant competitiveadvantages, including more loyal and valuable customers. Companiesthat fail to adopt predictive marketing are at risk of falling behind.The key is to start small and grow your efforts over time. Find a quickwin that can deliver immediate ROI.

CHAPTER 2

An Easy Primer toPredictive Analyticsfor Marketers

A natural nutrition company has been in business for 60 years. While itis an $800 million business, the marketing team is small and the com-

pany does not have an in-house data science team nor have the systemsto leverage the data it collected over the years about its customers, manyof which were very loyal. The company is using cloud-based predictivemarketing software to organize, understand, and utilize its customer datawith significant results.

The software discovered customers that make their first purchasewithout any free membership promotions spend 76.5 percent more thancustomers who join with a free membership offer. It also found thatcustomers who placed auto ship orders were likely to spend three timesmore over their lifetime. The company was then able to develop smartcampaigns around membership programs to encourage customers to signup and enroll in auto ship when possible.

The software also sorted its customer base into clusters, or groupsof customers with similar interests. In doing this, it found a weight-losscluster that was underserved and sent additional personalized communi-cations, resulting in a 300 percent increase in revenue. It became clear

23


that customers tend to make the majority of purchases within specificnutritional categories and rarely migrate across categories, promptingan initiative to focus communications to each cluster around the mostcomplimentary products for each cluster.

Lastly, the software came with a number of quick-win campaigns,triggered by specific customer behaviors such as abandoned cartand replenishment programs, which resulted in 28 percent incrementalweb conversions and 22 percent new orders from email reminders,respectively.

The nutrition business story illustrates how everyday marketers cannow use predictive marketing methods without ever hiring a singledata scientist. The company owns its own customer data, but relies oncloud-based software for predictive algorithms, advanced segmentation,and life cycle campaign templates. Even if they don’t have in-house datascientists or develop algorithms in-house, many marketers are curiousabout what is happening “under the hood” of such predictive marketingsoftware. This chapter is written for those marketers.

This chapter aims to give an easy primer on predictive analytics soyou will understand how predictive marketing software works. Thinkabout it this way: in order to use a word processor, you don’t need tolearn to program a computer. However, in the early days of personalcomputers, people took basic programming classes anyway before usinga word processor “just in case.” Similarly, this chapter is aiming toteach you the basics of predictive analytics algorithms. It will hopefullygive you more confidence using its outputs. You can safely ignore thischapter and jump to Part Two of this book. Knowledge about predictiveanalytics is not necessary in order to practice predictive marketing.

What Is Predictive Analytics?

Predictive models are used in many areas of business and everyday lifeincluding politics, fraud detection, or risk modeling, such as when cal-culating your credit score. For the purpose of marketing, we are lookingto use advanced math in order to predict individual customer behaviorand to group customers into the most actionable and meaningful ways.For example, using predictive analytics, you can predict whether andwhen a customer may be planning to buy next. You might also be ableto detect distinct groups of buyers in your customer data, such as cus-tomers who only ever buy on discount—so-called discount junkies—or

An Easy Primer to Predictive Analytics for Marketers 25

customers who buy a lot but return most items they buy—the returna-holics. Last, using predictive analytics you could predict what specificproduct a customer may buy next and recommend these products toyour customers proactively.

There are three types of predictive analytics that marketers shouldknow about:

1. Unsupervised learning (such as clustering models): Unsupervised learningfinds hidden patterns in data, without explicitly trying to estimate orpredict an outcome. For example, finding similar customers within alarge group of customers, such as those who like long-distance run-ning versus skiing, without explicitly knowing what groups exist orwho belongs to them. Unsupervised algorithms such as clusteringare thus typically used to unveil the true underlying segmentation ofyour data.

2. Supervised learning (such as propensity models or predictions): Supervisedlearning is used to estimate an output given an input, by training itwith sample inputs and target. An example is to estimate the lifetimevalue of a customer, the likelihood a customer will engage with yourbrand, or a specific product a customer might want to buy next.

3. Reinforcement learning (more commonly used for recommendations): Rein-forcement learning allows us to leverage hidden patterns and similar-ities in the data to accurately predict the best next steps, outcomes,products, or content for the user or a given event. Unlike super-vised learning, reinforcement learning algorithms are not provided atraining input/output sample but learn from a trial-and-error-basedlearning scheme.

Unsupervised Learning: Clustering Models

Unsupervised learning is about recognizing patterns in the data withoutknowing ahead of time what you are looking for or using explicit labels.One of the approaches is called clustering. For example, looking at thebuying behavior of customers, there could be a cluster of people whousually buy only when they are being given a discount. This group ofcustomers could be detected without imposing any specific pattern orhypothesis or knowledge of the data ahead of time, but rather the patternof this group would emerge by trying to group customers who are most“behaviorally similar” together.


The difference between clustering and segmentation

If segmentation is the process of manually putting customers into groupsbased on similarities, then clustering is the automated/statistically rig-orous process of finding similarities in customers so that they can begrouped. When you segment you know whom to target ahead of time;when you cluster you discover whom to target. Clustering is a method toautomatically discover segments in your customer base by using alreadyknown factors about your customers. Clustering algorithms, such ask-means and apriori algorithms, can analyze hundreds of customerattributes and previous customer interactions to reveal insights into cus-tomer behaviors and the forces driving those behaviors. This is differentfrom customer segmentation in the sense that most segmentation utilizesone or two factors, such as age or income in non-statistical ways to groupcustomers together. Also, as Swedish statistician Hans Rosling put it,“the problem is not ignorance, it is pre-conceived ideas.”

For example, if I am selling an expensive cocktail dress, I want tomarket to people who are most likely to buy the dress. So first I define thelimits of the target group: women with annual incomes over $100,000.Identifying and grouping customers who are women and have a highincome is the process of segmentation. I make an assumption that peopleoutside this segment most likely would not want a $1,000 dress. I maynot have information about household income but can probably estimateincome by looking at a customers’ zip code.

Clustering will help you discover which women might be most likelyto buy your new cocktail dress. Clustering algorithms look at manymore dimensions than just zip code. After looking at many variables,such as age, location, time of purchase, similar items purchases, and soforth, a clustering algorithm automatically groups customers with similarbehavior. For example, you might discover that it is women of a certainage who buy in the first two months of the year, who are the most likelyto buy a high-end dresses (including a cocktail dress) and that householdincome has very little or nothing to do with it. It is this automatic dis-covery of customer attributes that matter, and the grouping of customerswith similar attributes that we call clustering. Figure 2.1 demonstratesthe principle of clustering visually.

In this example we only looked at three dimensions—income, age,and time of purchase. In real life, the power of algorithms is that hundreds


Figure 2.1 The Principle of Clustering

of customer attributes can be analyzed automatically, until the algorithmsfind the attributes that are significant in teasing apart distinct groupswithin the customer base. Marketers now have hundreds of characteris-tics they can look at, such as brand preference, discount preference, timespent onsite, browsing behavior, length of call. It is just not feasible for aperson to go through hundreds of types of data to find the relationshipsbetween each variable, but for today’s powerful computers and softwarealgorithms this is a piece of cake.

Usually about 8 to 15 attributes together describe a customer cluster.You could consider this an automatically discovered persona, who youcan now start to market to. You might find that you have a statisticallysignificant group of customers, all young women, who buy every yearlike clockwork in February, and will only buy designer dresses that areon sale at that time—over the Internet. There might be another groupof women who are older, who only buy from you in the store, alwaysat full price, about every two months or so—but don’t ever go for thecocktail dresses, only for the casual wear. You get the picture.

Traditional methods often rely on human intuition and guesswork.Clustering, on the other hand, uses what’s called machine-learning algo-rithms to create customer segments. This allows computers to quicklystudy massive amounts of past examples and then learn from the


previously collected data to distinguish one customer group fromanother, find correlations a person might not have looked for, and leadto surprising results that marketers might not have discovered.

For example, let’s say you are an online retailer of fast fashion. It is dif-ficult to make specific product recommendations because your inventorychanges so quickly. You can, however, cluster your customers aroundproduct types in order to discover distinct buying personas. Perhaps youwill discover that certain types of products are often bought together, andthat some customers are one-category buyers, whereas others tend toshop across categories. Specifically, you might discover that some peoplewho buy active wear also tend to buy sunglasses, whereas another clusterof customers buy sunglasses together with beachwear. Using other cus-tomer attributes, such as location, gender, and time of purchase, you candistinguish these two clusters and market to them accordingly. One clus-ter might constitute women getting ready for vacation, whereas theother consists of avid runners—both men and women. Both segmentsare best addressed using different creative and content marketing strate-gies, but if you just look at the purchase of sunglasses you might missthe nuances.

We will learn how to use clusters in marketing campaigns inChapter 6.

Supervised Learning: Propensity Models

Propensity models, also called likelihood models, are what most peoplethink of when they hear the term predictive analytics. In mathemati-cal terms, these models use algorithms such as neural networks, logisticregression, random forest, and regression trees. The names of these algo-rithms are of little importance to marketers. What is important to knowis that propensity models make true predictions about a customer’s futurebehavior by learning from examples from the past. Examples include thelikelihood of a customer to buy a product or the likelihood of a prospectto engage with a website.

Propensity models are used heavily in direct mail and are oftenreferred to as response models, because these models predict a customer’sresponse, for instance to buy or not to buy, as a result of receiving a pieceof direct mail. Propensity models are guided-learning models, whichmeans that it takes some time to train the data and that the models


get better over time. As the model observes the actual outcome of theprediction, such as did the customer buy or not, then the model canadjust its algorithm and become more accurate over time. Therefore,most propensity models require a short training period and a testingperiod before you rely on the predictions entirely. You can acceleratethe training period by providing a historical data set as a training data set.

Figure 2.2 demonstrates how you need either a training and test-ing period or historical data to start making your first prediction usingpropensity models.

How to Use Propensity Deciles

A key term when it comes to propensity models is deciles. Instead ofusing individual customer scores, most practitioners group customersinto deciles, top 10 percent, next 10 percent, and all the way to thebottom 10 percent.

For instance, to predict how much money a customer would spendthroughout her lifetime, we would use what’s called a predicted lifetimevalue model. The top 10 percent of customers here might have an averagepredicted lifetime value of $1,000, whereas the average predicted lifetimevalue for the bottom 10 percent is only $5.

The decile approach is useful in two ways. First of all, it provides theaverage value for the expected behavior such as lifetime value or spend.Secondly, it essentially provides a rank order of your customers from themost valuable to the least valuable, or from the most likely to buy to theleast likely to buy, in 10 equal buckets. You could use this informationin many ways, such as to decide to whom to send an expensive catalog

Figure 2.2 Training Propensity Models


or to design a/b lift tests. Catalogs would only go out to the customersin the top one or two deciles or you can test what the lift is in send-ing catalogs to each decile, and determine the efficacy and effectiveness.Or you could use this model to decide who to invite to the upcomingfashion show of your new sneaker line. It is more advantageous to havecustomers there who will actually buy something after the show.

Similarly you could decide on different strategies for something assimple as an abandoned browse campaign: prospects who visited yourwebsite but didn’t convert are an important missed opportunity andmany retargeting solutions are available to follow these non-buyers acrossthe web. What if you could differentiate the offer you are making to thesecustomers based on their likelihood to buy? For people with a very highlikelihood to buy, a simple reminder might be enough to get them toopen their wallets, whereas for people with a very low likelihood to buyyou might offer them a discount or free shipping.

Figure 2.3 shows you the predicted customer response to a directmail campaign. This model is essentially predicting that the top 10 per-cent of customers, or the first decile, will make up 52 percent of allresponses to the direct mailer. Based on this you might decide only tosend this catalog to this top decile. If you are aiming for a certain numberof responses, say 70 percent, you may need to send this to the top twodeciles or you might consider sending to decile two and three because

Figure 2.3 An Example of a Direct Mail Response Model


decile one is likely to buy anyway, were as deciles two and three mightneed the catalogue to incent them to make a purchase. Or if you havea limited number of catalogs, or a limited budget, you can now aimthose catalogs at the customers most likely to respond. The other wayto use the model is to create a/b tests within each decile to measure thelift to see which decile can justify the cost of incremental catalogs withincremental profits.

This type of model can be used to predict future prospect or cus-tomer behavior. For example, from the moment I buy my first Guccihandbag, the luxury manufacturer can predict with a high degree ofaccuracy how many more handbags I will buy in the future. I may haveno intent to ever buy from Gucci again, but the brand knows better.Ironically, by comparing my purchases, website visits, and email clicks,as well as my age, sex, and location, to the behavior and demograph-ics of thousands of other customers who have come before me, Gucci’salgorithms can predict my future purchases better than I can myself.

Comparing Propensity Models and RFM Modeling

Before predictive analytics became widely available, the industry standardfor identifying who is likely to buy from you was a model called RFM(Recency, Frequency, Monetary Value). However, it has limited utilityand can be surprisingly hard to use in real life. Furthermore, even thoughRFM is often classified as a predictive model, it is nothing but a mereheuristic (using a rule of thumb, an educated guess) approach with nostatistical nor predictive foundation.

The idea is that if a customer purchased from you recently, frequently,or has spent a lot of money with you, he or she is likely to buy from youagain. There is certainly no argument that “how many days has it beensince the customer last bought from us?” (recency), “how many timeshas a customer bought from us?” (frequency), and “how much revenuehas the customer generated for us?” (monetary value) are all excellentvariables to try and predict whether the customer will return to makeanother purchase.

However, this technique has limitations. For one, it severely restrictshow companies use their own data. There are so many other variablesthat can be derived from data that can serve as additional excellent pre-dictors. Also, the old saying, “past results are no guarantee for future


performance,” holds true for RFM modeling as well. RFM is exclusivelylooking backward, rather than comparing customers’ current behaviorto the future behavior of others who came before them. Using RFMyou will not be able to recognize high-value customers before they havebought from you. You might also ignore past high-value customers whohave defected to the competition.

Let’s look at an example: there might be a pattern where customersbuy a couple of times and then disappear. The only way to find this outis by comparing customer behavior to other customers just like themto make predictions about future behavior. If most customers buy threetimes and then disappear, then the likelihood of a customer to buy afterthree times is actually very low, whereas the RFM model would putthis customer in a “highly likely to buy” segment. Thus, good, respon-sive customers might end up in not so valuable segments and thus getignored for promotional mailings. The opposite might happen as well—not-so-great customers might end up in valuable segments. A responsemodel that ranks customers by their value as opposed to the value oftheir segment can take care of this problem.

Also, RFM models can only make predictions about the likelihoodto buy in an environment where there are frequent purchases, such asretail. RFM models do nothing to predict a customer’s lifetime value,a customer’s likelihood to unsubscribe from a service, such as a sub-scription service or your email list, or the likelihood of a web browser(nonbuyer) to convert into a first-time buyer. Supervised learning mod-els can be used to make predictions about all of these customer behaviorsand more. In many comparison tests, where the top 50 percent of cus-tomers are selected with RFM and propensity models, propensity modelsare on average 40% more accurate than RFM. This in many cases trans-lates to 20–25% less promotional costs.

Ironically, propensity models are not only more accurate, they arealso much easier to use for the everyday marketer. Rather than havingto choose which combination of potentially hundreds of possible com-binations of recency, frequency, and monetary value to use for acampaign, you can just decide which of the ten propensity deciles toinclude. With predictive analytics, marketers are presented with a list thatautomatically ranks customer from the most likely to buy to the customerleast likely to buy.

We come back to some specific supervised learning models and howyou can use them in Chapters 7 and 8.


Reinforcement Learning andCollaborative Filtering

Reinforcement learning is typically used in combination with collab-orative filtering models. The common marketing application for col-laborative filtering models is recommendations. From a technical point ofview, recommendation models make use of the latest machine-learningtheories in the field of collaborative filtering, bayesian networks, and fre-quent item sets. Time-decay functions are used to take into account thefact that recent behavior has more predictive weight than older behavior.Finally, reinforcement learning is applied to “educate” the model to thecustomer’s preferences. Again, we just mention these names in case youwant to dig deeper.

Collaborative filtering models can recommend products, content,or just about anything else. These recommendation models were madefamous by Amazon with their “customers who liked this product, alsoliked…” suggestions. Recommendation models are a fantastic way togrow the value and retain your customers, by suggesting relevant prod-ucts or content they will be interested in. Suggesting relevant prod-ucts will drive revenue directly, whereas suggesting relevant content willincrease the engagement with your brand, and indirectly create morehappy and loyal customers.

It is important to serve up recommendations that fit the contextof where they are presented to customers. Bad or out-of-context re-commendations will be viewed as “creepy,” “intrusive,” or “irrelevant.”Recommendations should come at the right time: for example, just asyou are about to check out your online shopping cart it makes senseto receive a “customers who bought this, also added…” type of recom-mendation. Perhaps two days after, it is appropriate to receive a thank youemail suggesting useful follow-on purchases. Those who bought a woodgrill might now be recommended a cookbook or refill wood chips. It isbest also to expose this context to the consumer. The more transparentyou are, the more consumers will accept and act on your recommen-dations. Companies that successfully utilize recommendations are nowstarting to offer explanations of their recommendations to remove theintrusive nature. You will see Amazon use words such as “because youlooked at this product, you might also be interested in these products,”or “people who bought this product, bought it together with this otherproduct.” These brands are also starting to give consumers control over


which products are recommended. You can go to a preference center andexclude certain items from being considered for the recommendationalgorithms.

If you have worked with recommendations before you know thatthere are many details that matter when it comes to recommendationalgorithms. First, to stay relevant, the recommendations are ideallyrefreshed in real time, and the underlying models are refreshed daily foreach customer to take into account a customer’s recent behavior butalso the change in behavior of other similar people.

Also, you want to make sure that the model does not recommendproducts that are out of stock or products that have a high return rateor bad reviews. Some retailers do not recommend items that are on salebut opt for higher margin items. A good recommendation model alsoallows for marketers to manually enter merchandising rules that tweakthe algorithm.

Different Types of Recommendation Models

There are three good use cases for recommendation models: upsellrecommendations, next sell recommendations, and cross-sell recom-mendations. Each of these has a different place in a marketer’s arsenal.Also, recommendations can be made for “products that are generallybought together” or can be made specifically to a person based on pastbehavior. Let’s start by explaining the difference between those two.

Products Generally Bought Together Products that are typicallybought together are not customer-specific. At our company, we some-times call these product-to-product recommendations, though this is not acommon industry term. These are the type of recommendations that aredisplayed on a product page, and will recommend other relevant prod-ucts to all visitors of that product page. Products that are typically boughttogether are called product-to-product recommendations. They answerthe question: “Customers who bought this product, typically alsobought what other products?” In this scenario, two people browsingthe same product will receive the same recommendations. In Figure 2.4two people browsing the first bikini set will receive a recommendationfor another bikini, a swim cover-up, and a tote bag. These types ofgeneral recommendations are especially relevant when you don’t know


Figure 2.4 Products Typically Bought Together

much about a specific customer, such as an anonymous first-timebrowser coming to your website.

User Specific Recommendations If you have more informationabout a specific user, then you can go further than generic recommen-dations. Let’s say you know the person looking at this bikini set is amale and this is the first time that he is looking at swimwear. Normallyhe browses and purchases electronics from your website. While he isbrowsing, presumably for a gift or something, it might be appropriate torecommend other bikini tops. However, when sending the customer athank you email two days after his purchase with recommendations ofwhat to buy next, you would be better off making recommendationsspecific to this person that take into account the entire history, not justthe most recent browsing session.

These are product recommendations specific to a given customer.In our company we sometimes call these types of recommendationsproduct-to-user type recommendations. You can replace “product” with“content” or “person” or whatever you are trying to recommend. In thiscase, two users looking at the same product at your site would receivetotally different recommendations.

User-specific recommendations are not limited to physical goods.You can replace the word “product” with “content” or “event.”


For example, Shazam is recording the music tastes of consumerswhen you tag (Shazam) songs you like. Based on your personal tastesShazam will recommend concerts that are of interest to you, but onlyif the concerts actually take place in your geographical area. For theserecommendations to be successful, Shazam needs to know not only thetastes of its customers but also their physical locations.

The Predictive Analytics Process

Let’s walk through the different steps a data scientist, or analytics soft-ware, goes through to make accurate predictions or recommendations.Most of what we describe here happens under the hood and marketersdon’t need to worry about any of this. Figure 2.5 gives an overviewabout what is happening under the hood, either of out-of-the-boxpredictive analytics software, or the steps your internal data scientistswill have to go through if you are building your own predictive analyticsmodels. We don’t mean to scare you by being very specific about someof these steps in this chapter. However, we do want you to realizethat the do-it-yourself route to predictive marketing requires highlytrained data scientists. There is a lot that comes with developing anddeploying predictive algorithms for marketing, and if you are startingnew and want our advice, we strongly recommend that you use anoff-the-shelf software package suitable for your industry to take careof the steps described in this chapter automatically. When you exhaustoff-the-shelf models, and you have the budget and need for a datascientist, you can easily evaluate the cost-benefit equation with the

Figure 2.5 Overview of the Predictive Analytics Process


experience you gained (and you’ll find it easier to convince the CFOto make the incremental investment).

Data Collection, Cleansing, and Preparation

Data cleansing and preparation is the most important and most ignoredstage in predictive analytics. In some cases, the data might be missing orincorrect as collected. Data cleansing is used to correct things like namesand addresses to make sure the computer will know that a customerlives in California when her state is listed as CA. We discuss the processof collecting customer data and linking it together to form individualcustomer profiles in detail in Chapter 3. However, even after building360 customer profiles, there still is significant work to be done gettingyour data ready for analysis. Not all data collected is immediately usableand the results could be skewed by missing data or outliers, data mea-surements that are either too low, too high, or do not fit the underlyingdata-generating system.

If you are considering building your own predictive analytics capabil-ities, make sure you address who will do the data collection, integration,cleansing, and preparation. Chances are, your typical data scientist is notgoing to be satisfied doing this work and is going to expect that you hirea separate software or integration engineer to do this work.

Outlier Detection

Outlier detection often makes a big difference in the accuracy of predic-tive models. For example, if a customer at an electronics retailer came inand bought 50 televisions for $50,000 when the average customer at theretailer spends $500, this high-spender will skew the average order valuemetric. In electronics retailing, these types of outliers where there are fewusers making large purchases are indeed quite common. People makingsuch large purchases could be middlemen who are buying items liketelevisions to take out of the country and resell. These are not normalconsumer customers, but rather gray market resellers. If this situationwasn’t recognized and corrected, the retailer would think these are greatVIP customers. Not recognizing this creates two problems: distorting thedefinition of VIP customers so the true VIP customers would be left outand masking an opportunity to market to this group of resellers in a moreprofitable way.


To correct for the outlier, your data analyst or predictive marketingsoftware will need to detect and either remove the outlier or replaceit with a number at the high end of the distribution (e.g., the lowestspend of the top 10 percent customers is $2,400, so replace the $50,000with $2,400). This replacement is only done for modeling purposes.Alternatively, you can treat these customers as a separate group altogetherand create specialty programs for this one segment.

In another example, one retailer was measuring foot traffic at eachstore but would miss out on data for certain days whenever the measur-ing device would get knocked off by cleaning crew. To correct for themissing data, the retailer applied an imputation based on the three-weekaverage for the same days in the week as the missing days. Imputation isthe art and science of replacing wrong or missing information. Depend-ing on the specific data elements, there are various techniques for this:

• Replace with static or temporal averages.• Model the data based on other variables available. For example, you

can model a vitamin store customer’s age based on whether she buysvitamins geared toward women above age 50.

• Random selection from the underlying distribution. For example,if the foot traffic data is missing and this data usually follows abell curve, then randomly generate a number from the underlyingdistribution.

Imputation is a great way to make up for missing data until the problemis corrected at the source.

Another example of imputation is asking customers for their birth-days. This is a great piece of information for modeling and action pur-poses but not all customers want to provide this information. In suchcases, the predictive model would either discard the birthday as an inputor discard customers with no birthday.

Feature Generation and Extraction

Once your data scientist or predictive marketing software has cleansedthe data for missing information and outliers, there are two other factorsto consider: (1) the data may be too large to use as is, or (2) data in itscurrent representation may not be suitable for the models. Feature gen-eration and extraction deals with turning data into information that the


models can digest and throwing away unnecessary or redundant infor-mation. Think of feature generation and extraction as separating the sig-nal from the noise. Feature extraction deals with removing unnecessaryinformation by either throwing it away or transforming it to eliminatethe noise. There are quite a few mathematical methods to use, but theshort explanation is to utilize algorithms to be able to extract the maxi-mum amount of information from the data, regardless of what you willuse it for later on. This optimal extraction leads to less noisy data, henceincreasing the accuracy of predictive analytics.

There are tricks you can use to make your data easier to use.For example, when trying to analyze the number of orders from acustomer you can look at the numbers in absolute terms or you couldtake the logarithm of the number, creating a new variable where if a cus-tomer has 1 order or 10 orders, it’s the same difference between having10 orders and 100 orders. It is a simple transformation that can have apowerful impact.

Another example could be taking the ratio of certain variablesinstead of using absolute numbers. For instance, instead of returnrevenue and shipped revenue per customer, you can calculate the ratioor percentage of revenue that comes through returns.

Classifier and System Design

The next stage in the process used by data scientists or predictive mar-keting software is choosing, architecting, and fine-tuning the correctalgorithm. In machine learning, there are two important concepts thatneed to be understood. One is the no free lunch theorem, which states thatthere is no inherently better algorithm for all problems out there. This isimportant to understand so the data scientist chooses the right algorithmfor the right problem, and not use the same algorithm for every problem.The other concept is called the bias-variance dilemma, which states that ifyou go deep in developing an approach and algorithm to solve a specificproblem, then the system that is biased toward this specific problem getsworse and worse performance in solving “other” problems out there.The lesson learned here is to understand that no algorithm is inherentlybetter than the other. If you develop your own algorithms it means youprobably have to develop multiple algorithms for multiple situations. Ifyou buy off-the-shelf predictive marketing software, you want to make


sure to choose a vendor that focuses on your specific vertical and businessproblems, and/or to choose a vendor that has self-learning algorithmsthat can adjust to your specific situation automatically. Correctly archi-tected software solutions usually have multiple models competing witheach other, and the “champion” is selected against “challengers” that isunique to the customer’s domain of data. This maximizes performanceand eliminates the need for hand-coded custom models.

Just writing an algorithm is not enough when it comes to predictiveanalytics. Before you can start to use an algorithm, you need to back-test that it actually works. If you use a predictive marketing softwarepackage, your vendor will have done this for you already. However, ifyou are developing your own predictive analytics models in-house youwill have to worry about training, testing, and validating your modelsbefore you can start to use them. The time needed to develop predictivealgorithms can be divided into 80 percent training, 10 percent testing,and 10 percent validation. This means that after writing the algorithm,data scientists need to spend considerable time training and testing thealgorithm to make sure it works accurately.

For example, when developing a likelihood to buy model, if 1 per-cent of 10 million customers buy within the next 30 days, then fortraining, we use 100,000 customers who bought in the past month andrandomly select 100,000 customers who didn’t buy anything in the pastmonth, so that the total dataset has 200,000 customers in which 50 per-cent bought and 50 percent didn’t. This oversampling produces betterresults, because it focuses the model to detect between potential buyersand non-buyers.

The Last Mile Problem of Predictive Analytics

Most data scientists don’t worry how marketers will use their predic-tions. Frankly, most data scientists don’t know enough about marketingand marketing systems to embed predictions into the daily routine ofmarketers. An email marketer at a large national department store oncetold us: “Brides register with us on our website and leave a lot of personalinformation. That information is in our customer data warehouse some-where and we probably even analyze it. However, as an email-marketingmanager, I am unable to run a simple campaign that takes into account


some of the preferences or dates that the bride has shared with us.”We call this the last mile problem of predictive analytics.

Especially in organizations with in-house data scientists, the out-comes from predictive models are often not easily digestible or usable bymarketers. It is often very difficult for marketers to put predictive analyt-ics into action—to connect the dots from analytics to the daily campaignmanagement of email, web, social, mobile, direct mail, store marketing,and customer interactions in the call center.

For customer predictions to be profitable, predictions need to beput in the hands of all the customer-facing personnel in your organiza-tion. If you can’t surface recommendations to the personnel in your callcenter, the upsell might never happen. If you can’t use likelihood to buysegments to decide whether to send an abandoned cart holder a discountor a reminder, you are leaving a lot of profit on the table. In Chapter 14we look at technologies that bridge this last mile problem.

Now that you have a basic understanding of predictive models andhow they can be used for marketing, we get to work. In the rest of thisbook we give you more examples of each of the three types of machine-learning models discussed in this chapter. We cover how to use eachof these models for marketing in much detail. We also propose how tooperationalize each of the models, connect the findings to specific mar-keting campaigns that can drive immediate revenues or profits for yourorganization. Without marketing actions, there may be analysis paralysis,thus no customer delight nor incremental profits.

CHAPTER 3

Get to Know YourCustomers First: BuildComplete CustomerProfiles

Before you can leverage customer data to find new growth opportuni-ties or hypertarget your customers, you first need to aggregate, clean,

and analyze that data. This is no easy task. If yours is like most compa-nies, customer data is all over the place, full of errors and duplicates andinaccessible to everyday marketers. Fortunately, predictive technologycan help clean up your data mess. Your information technology (IT)department and potentially outside vendors can also help.

Bosch understood that its customers were no longer only homeimprovement retailers like Home Depot and Lowe’s. Instead, the rise ofonline shopping allowed their end customers to go directly to its websiteand engage with the brand, without ever stepping foot into a big boxstore. That created a challenge for Bosch, which had limited insight intothe end-user data of who was buying their products.

Bosch realized it had to do a better job marketing directly to its endcustomers, but to do so it needed to get a better handle on its customerdata. Like many companies, Bosch did not have an internal data teamand was relying on an outside service provider. That meant that for every

43


question about customers, it had to ask the external provider to createa new report. So Bosch decided to bring its customer data in-houseby using modern cloud-based software, rather than to rely on an outsideconsulting firm. Now, the company can aggregate customer informationfrom various data sources and link it together into individual customerprofiles. It tracks purchases, but also newsletter signups and product reg-istrations. Bosch has learned that even without using customer data forone-to-one personalization, the 360-degree view of customers has madeall the difference when making decisions about how to market to its cus-tomers. The next step is to understand how this data can be utilized tosupport marketing activities with partners like Home Depot and Lowes.

A New York boutique fashion brand had the same experience.Once it collected all customer data into a single profile, it was ableto give customer service personnel access to these records. Now if acustomer writes or calls in to complain, it could look up if this customerwas a high-value customer or not, and get more context overall to havea better conversation with this customer, and respond appropriately tothe complaint.

The Bosch and fashion brand examples illustrate clearly that evenwithout advanced segmentation or analytics, organizing customerinformation across all customer touch points, including web, transac-tions, mobile, email, store, call center, into a 360-degree profile foreach customer can be a huge win. The Bosch example also illustratesthat customer data is not only for direct-to-consumer marketers.Of course, for those companies that have wholesale distribution,getting consumer data can be a challenge. Industry standard is suchthat wholesale partners do not pass along customer data. However,brands can develop a program whereby they utilize a registrationor warranty program where they put a card in the packaging forconsumers to sign up online to register online. The card will ofteninstruct you to go to a specific registration URL with the brandin exchange for registering the product for future issues with theproduct or simply register their email address to get the latest productinformation and care instructions directly from the brand. Surprisingly,many consumers are will to do this, and though you may not have allthe transactions associated with the purchase, you’ll have their emailor mailing address. Service records can serve the same purpose. Awell-known European retailer of white goods is using repair records

Get to Know Your Customers First: Build Complete Customer Profiles 45

as triggers for personalized marketing campaigns. Lastly, brands sellingthrough wholesale may also consider using visits to their corporatewebsite as a proxy for product interest. Clusters can be built based onbrowsing behavior alone.

This chapter helps you understand what data you need to collect first,if you, too, want to bring customer data in-house as a core competency,and how to prepare your data for analysis. We will also add a few words onhow to partner successfully with your IT counterparts to bring customerdata projects to completion.

How Much Data to Collect

In this book, we focus on customer data provided to a companydirectly by its customers, not data derived or purchased from thirdparties. Marketers have more historical and real-time data on customersthan ever before. This type of data is also called first-party data,data that is owned by the marketer that reflects direct interactionsof the consumer with the brand and has the richest informationcompared to third-party data sources. Third-party data sources needsto be anonymized and sometimes come close to breaching privacyboundaries. We talk about this in more detail in later chapters. Eachindividual consumer generates hundreds of data points each day, whichwhen multiplied by thousands or even millions of customers result intruly big customer data. Data is said to be “big” when there is a lotof volume, variety, and velocity. For customer data this is certainlytrue. In fact, the data most companies collect from their customershas grown so large and varied that no human being can analyze itanymore without the help of computers and software.

Marketers who can find a way to harness the power of all this cus-tomer data will have a significant competitive advantage. Figure 3.1outlines some design principles that may be useful when deciding whatdata to collect. The most important design principle is to start with theend in mind. Often, marketers make data collection and integration aproject on its own without specifically outlining what the data will beused for and how it will be maintained to ensure it is in great shape.

When considering how much customer data to collect, also takeinto consideration how valuable that data is, as well as how easy itis get. This may vary by company, but remember not to overdo it.


Design Principle forData Collection Example

Frequency How frequently to collect data, and on which event triggers?Derived data Derived data are implied data elements. A customer who visited

the website and browsed a product five times and each timebought from a store in the following seven days could belabeled as a customer who collects information online, butshops offline.

Granularity Web data could be collected click by click or in some casesa summary about the web sessions could be enough.

Insights to be derived If the goal is to predict customer upside potential, the type ofproducts a customer purchases is important, as well as the zipcode the customer lives in. The Insights to be deriveddetermine which data we collect.

Actionability Data collected should be actionable directly or indirectly.Collecting the sports interests of customers is actionable for asports retailer, but not for a company that does tax consulting.

Accuracy When asked for age, quite a few customers randomly typeanswers, more often in cases where marketers use it for gatingcontent or a sign up. Marketers need to deal with theseinaccuracies through imputations. Imputations is the processof replacing missing values with substitute values.

Fill rates Marketers often want to collect data on customers usingprogressive profiling in order to boost fill rates.

Storage How much or how long to keep the data depends on the“currency” of data. Web browsing data is often not relevantafter a few weeks, whereas purchases stay relevant for years.

Accessibility Data collected should be accessible to marketers for analysis andaction. Too often customer data is stuck in silos, inaccessibleto everyday marketers.

Figure 3.1 Design Principles for Data Collection

Many marketers obsess over collecting all customer data at once. Thatis a big mistake. It is very easy to get caught up in a long and slow dataintegration project without ever seeing results. Your goal should be tocollect just enough data to find new growth opportunities and startmarketing programs that deliver results. You will be surprised by howlittle information you need to get started. Once you can show resultsfrom your initial data-driven campaign, it will become easier to get thecooperation from other departments, such as your internal IT team, tocollect more customer data.


Big data inherently has a lot of noise in it. So data collection needs tobe paired with techniques to find the signal in the noise. The importantpoint here is to know how to extract the information from this largedataset to make it manageable, insightful, and actionable. Figure 3.1 givesan overview of some of the questions marketers need to answer, in col-laboration with their technology teams, about the data to be collected,integrated, and analyzed.

What Type of Data to Collect

Traditionally, marketers have mainly used purchase data and customerdemographic data. These days, marketers also have access to more behav-ioral data points, which brings temporal information. This temporalinformation can be used to derive context and to make marketing morerelevant in time. When we interact with a company or a brand, each ofour actions leaves a digital footprint, which gets recorded in a database.For example, for every purchase that somebody makes online, we nowhave about 50 data elements on each customer before that purchaseis even made. These include what customers clicked on in an email,whether they clicked on Google Adwords, what reviews they may haveleft, activity on social networks, complaints, and calls to customer sup-port centers. The amount of behavioral data available has exploded inrecent years. It is easy to get overwhelmed, but not to worry. We showyou where to start.

While no two businesses are alike, Figure 3.2 gives you an example ofwhat a phased data integration strategy might look like. In this example

Phase 1 Phase 2 Phase 3

BehavioralPurchasesWeb behaviorEmail behavior

BehavioralCall center interactionReturns and complaintsCustomer meeting notes

BehavioralSocial interactionsReviews and surveysLoyalty program interaction

DemographicHousehold affiliationAccount groupingLocation

DemographicGenderU.S. census dataVertical and size

DemographicAdditional third-party data

Figure 3.2 Three Steps to Customer Data Collection


the assumption is your end goal is to drive customer engagement andincrease customer lifetime value.

High-priority behavioral data for both consumer and businessmarketers includes purchases, web visits and email clicks. High-prioritydemographic data for consumer marketers could include the gender,age, and location of a customer, and for business marketers, more likelythe industry, size of the organization, title of the buyer, and the locationof the buyer’s headquarters. Purchases alone will already give you awealth of information. In fact, every single purchase generates lots ofinteresting meta-data points, such as time and location of the purchase,product purchased, and the sales person involved in the transaction.Figure 3.3 summarizes more of these points. We will describe each ofthe different data types and how to collect them in detail in Appendix A.

Data You Can Collect at the Time of Purchase

Time of the purchaseDate of the purchaseDate of shipmentBilling address for the purchaseShipping address for the purchaseName of the buyerGender of the buyer (derived from the name)Shipping revenuesHow long ago was this purchaseChannel of the purchase (for example, online or offline)Product that was purchasedProduct category that was purchasedBrand that was purchasedSalesperson involved in the purchase (B2B and B2C)Price of the purchaseDiscount applied to the purchaseRevenues generated by the purchaseCost of goods sold for the purchaseMargin of the purchaseTax collected on the purchaseShipping revenues associated with the purchaseWhether this was the first purchase or a repeat orderNumber of products in the orderTypes of products included in the orderWhat type of device the customer used to make the purchase

Figure 3.3 The Anatomy of a Purchase


A special challenge exists when it comes to in-store or otherin-person purchases. Many store purchases are anonymous. You mighttry to collect email addresses in the store by offering customers anelectronic receipt or an incentive in exchange for an in-store newslet-ter signup. You might also give store associates a discount or othercompensation to collect email addresses.

For a well-run program, expect in-store capture rates above 60% andas high as 95%. Improving data capture rates can dramatically improvethe success of in-store clienteling. Clienteling is a technique used byretail sales associates to establish long-term relationships with key cus-tomers based on data about their preferences, behaviors, and purchases.Clienteling is intended to guide associates to provide more personal andinformed customer service that may influence customer behavior relatedto shopping frequency, lift in average transaction value, and other retailkey performance indicators. From the customer’s perspective, clientel-ing “could add a layer of personal touch” to the shopping experience.Clienteling with big data dramatically improves upsells and consumersatisfaction because the consumer develops a relationship with a salesassociate that has a vested interest in securing their loyalty. Retentionrates improve dramatically for brands that make 360 customer profilesavailable to sales associates and that associate consumer transactions withindividual sales associates.

Mavi, the international jeans and Apparel Company from Chapter 2,wanted to tie POS transactions to customers to understand customersindividually. Mavi introduced a loyalty card program to accomplishthis. When it started the program, in its first year, only 20 percent oftransactions were tied to individual customer. By year four it had tiedclose to 90 percent of transactions to individuals. There are a few thingsMavi did right when introducing this program: First, it put goals andmeasurement in place and always focused on improving. Second, itdelighted customers with the data it collected, for example by sendingthem highly personalized offers to come back to the store and getextra points to buy again. By doing so, customers wanted to identifythemselves, because they saw the benefits. Third, the store personnelsaw the benefits and were always educated about why they did it. It wasa cultural focus and shift.


Loyalty program interaction itself is another important data point:the use of loyalty rewards can be very different from customer to cus-tomer. In fact, we have found that when grouping customers based ontheir behavior, the use of loyalty points is often a differentiating factorand some groups of customers— such as men—are more inclined to beswayed by reward offers than others.

In another example, Walmart created a mobile application calledSavings Catcher, which promised customers who downloaded the appthat Walmart would compare the prices of their purchases against pricesoffered by Walmart’s competitors and refund customers the differencewhen it found a lower price at another retail store. By providing anincentive to customers, Walmart was able to collect millions of emailaddresses and analyze information, like which products customersbought and what time they typically went shopping.

Business purchases are likely tracked in your enterprise resourcemanagement system or customer relationship management system.These systems are also a rich source of demographic information, suchas the location of the customer and the relationship between contactsand accounts, as well as between salespeople and accounts.

Preparing Your Data for Analysis

When it comes to your customer database the saying “garbage in, garbageout” holds true. If you base your customer segmentation or predictivemodels on bad or incomplete customer data profiles, you will get thewrong recommendations for your customers. Therefore, data wranglingis a huge part of the job. Data scientists will tell you that data preparationbefore analysis can make up 95 percent of all the work.

Without a single view of the customer, it is impossible to trulyunderstand any single customer or to draw any conclusions about cus-tomer trends. For example, if you can only see a person’s in-store pur-chases, but that person makes 90 percent of his or her purchases online,you may believe this is an unprofitable customer when in reality thisperson could be one of your VIPs. Similarly, if somebody frequentlybrowses your website but always ends up buying in the store, you maymistake the website customer as “low value.” In a different scenario, acustomer spends a large amount and buys often, but returns items just asfrequently or makes frequent calls into your call center. This customerlooks like a “high value” customer, but is in fact unprofitable.


Figure 3.4 The Data Preparation Process

This means that you need to be able to integrate, link, and dedu-plicate all the information that you have collected. This is not an easytask. Figure 3.4 gives you an overview of the data preparation process ata high level. We will go through each of the steps in the data preparationprocess briefly.

Cleaning and Validation of Names

After receiving the raw data, the first thing you want to do is to validatethe names, postal addresses, email addresses, and phone numbers in thecustomer files you received. Without this, software algorithms will beunable to link the right activities to the right data records. Examples ofcommon mistakes that need to be corrected before matching records toreal people include:

• Middle names or initials might be included or excluded: William Land William Louis might both be variations of the same person—William Morrison.

• Two-person contacts might not be matched unless corrected:“William & Cathy Morrison” should be matched to WilliamMorrison’s customer record.

• Replacing abbreviated names: Wm and Bill and William may all bethe same person.

• Removing honorifics: Rev. Bill Morrison and Dr. Bill should bothbe matched to William Morrison’s record.

• First and last name swapped: Bill Morrison and Morrison Bill arelikely the same person, especially if they live at the same address.

• Slight variations in name spelling, like Katie and Cathy.

There are many common mistakes that software can easily correct.Software can automatically normalize names, such as change Bill toWilliam or vice versa, and recognize and tag men and women.


Metaphonic algorithms are also correct words that have similar pronun-ciation, such as Katherine, Cathy, and so on. Software can automaticallystandardize names so that different records from the same customerwill now match and be linked to a single customer ID. For example,Michael is the same as Mike and James is the same as Jim, and soforth. As part of name standardization and verification, software canalso change the case of the name. So if the name was in all capitals(WILLIAM) or all lowercase (william) we make it have first letter capital,remainder lowercase (William). This may seem trivial, but softwarealgorithms are not people—and tend to take things literally. Withoutcorrection and normalization, these records would not be matched to thesame person.

Cleaning and Validation of Addresses

For mailing addresses, validation is important to make sure that eachpiece of expensive direct mail you are sending is actually deliverable.Address validation can reduce mailing costs by as much as 80 percent.These are some ways to validate mailing addresses:

• Canada and U.S. address coding according to USPS standards. Make surethat the address is complete and correctly written including a zip codewith a four-digit append to have the most accurate address possible.

• NCOA (National Change of Address). Each address in your database canbe checked against the NCOA database to make sure that the recipienthas not moved since you acquired his or her address.

• CASS certification. CASS certification is a requirement for all mailersin order to receive certain mailing rates from the USPS based on thequality of their addresses.

• DPV (Delivery Point Validation). This is the highest level of addressaccuracy checking—where each address gets checked against adata file to ensure that it exists as an active delivery point forthe USPS.

• Address type flag. If relevant, address parsing can also append an addresstype flag, such as whether the address is a residence or a business.

• MSA/region append. Based on each address, software can append thelongitude and latitude of the location, but can also match this againstadministrative regions.


Email address validation is equally important. Email verification ser-vices can improve client reach and reduce the danger of damaging yoursender reputation. Checks that can be performed automatically include:

• Syntax correction. Syntax correction in names and addresses automat-ically removes illegal characters and fixes host names. For example,software algorithms can automatically fix common domain name mis-spellings (gmai1.com = gmail.com, for example). Software could alsovalidate and correct illegal characters. For example, if it says gmail,cominstead of gmail.com, the correction would be to replace the commawith the period.

• Mail test. As part of validating email addresses, software can automati-cally “ping” the domain in the email address to ensure this domain isavailable as a mail exchange. Software can also automatically maintaina list of invalid emails.

• Invalid email filtering. Certain common default values, such [email protected], can automatically be detected and filtered out.

• Casing standardization. As with names, when it comes to checkingemails, software can automatically standardize all addresses to includeonly lowercase letters.

Linking and Deduplication

To eliminate duplicate copies of repeating data, you can use a techniquecalled deduplication. It is important because it can increase the accuracyof key performance indicators and metrics (like the lifetime value of acustomer). It also helps you avoid targeting the same person twice, whichdoesn’t look very professional and can be very expensive when it comesto campaigns like direct mail. The attributes of each of your contactsshould be merged according to a set of priority rules to obtain a mastercontact list. At this time, you should also associate the right contacts withthe right households or the right corporations.

In order to accurately deduplicate your data, you can use softwarealgorithms to do what’s called fuzzy matching. Fuzzy algorithms com-pute a similarity score between attributes such as Names or Addresses.When the similarity is higher than a defined threshold, the entities areconsidered duplicates. Hence fuzzy logic will “estimate” whether twocustomer names that are similar, but not exactly the same, might be

mailto:[email protected]


the same person. Some customer attributes, such as a customer’s homeaddress, will have a higher weighting in this estimation. William Morri-son and Bill Morrison are likely the same person if they live at the sameaddress, but unlikely the same person if they live in different states.

Now that all the hard work of collecting and cleaning customer datais done, in Figure 3.5 we give an example of the information that may beincluded in the profile of a single customer. Just with this profile alonethere is a lot of value. You could give your sales team, customer successteam, call center team, or in-store personnel access to these profiles andthey will surely be able to serve customers better with this type of infor-mation at their fingertips. In the next chapters we take the next stepsto actually using this information for customer analytics and to createunique and meaningful experiences for each and every customer.

It is not unusual to make important discoveries about your customersafter integrating all customer data. In the case of a jewelry company, theyfound that though 70% of the product line featured women’s jewelry,50% of the buyers were men buying jewelry as gifts. This group wasvirtually untapped until a demographic append was done to uncoverthis target. Even women purchasing jewelry were often shopping forpresents, such as heirlooms for daughters and sons for special occasionslike graduations from high school and college. Lastly, after collectingall customer data, this company found that there was very little overlapbetween metal types. Gold buyers continued to buy gold and silver buy-ers continued to buy silver. Of course all of these insights changed theirmarketing strategy entirely.

Working with IT on Data Integration

We highly recommend that you work with your information technol-ogy (IT) department when collecting and integrating data into a single,real-time view of the customer, so we want to conclude this chapterwith some advice for successfully partnering with your IT team:

Don’t go it alone. A study by the CMO Club found that 88% of mar-keting executives admit that projects run outside of IT control “some-times” (53%) or “often” (35%) run into problems. According to a 2014Accenture report, only one in 10 marketing and IT executives say col-laboration between the two departments is at the right level. CIOs arein a key position to guide technology initiatives across the organization,


General demographic• Name• Email• Gender• LinkedIn search• Address• Location (latitude and longitude)• Google Map view of Address

Predictive analytics

• Likelihood to buy (“high”)• Behavior-based cluster (“discount

junkie”)• Product-based cluster (“laptop buyer”)• Brand-based cluster (“Dell”)• Life cycle cluster (“new customer”)• Product recommendations

Contact strategy

• Preferred channel• Preferred store• Closest store• Preferred brand• Marketable by Phone? (Y/N)• Marketing by Mail? (Y/N)• Marketable by Email? (Y/N)

Life cycle cluster values could include

• Prospective customer• New customer• Repeat customer• Lapsed one-time customer• Lapsed repeat customer• Inactive one-time customer• Inactive repeat customer

Purchase analytics

• Lifetime revenue (for example,$2,007)

• Lifetime margin (for example, $576)• Lifetime order count• Last 12 months revenue• Last 12 months margin• Last 12 months order count• Average Order Value• Last 12 Months Revenue Segment

(for example “top percent ofcustomers”)

• Prior 12 Months Revenue Segment• Revenue Trend (Up, Neutral, or

Down)

Behavior

• Last order date• Last order channel• Last order revenue• Last web visit date• Web visit count• Last send date• Last open date• Last click date• Email open count last month• Email open count two months prior• Email click count last month• Email click count two months prior• First order date• First order channel• First order revenue• Distinct channels• Distinct products• Distinct categories• Last five orders (channel/date/product/

brand)• Last five on site search (search term, date)

Figure 3.5 Example of a Customer Profile


ensuring reliability, data privacy, security, and compatibility with the cor-porate technology stack amongst others. As a marketer you can play therole of “chief experience officer” but accept IT as a strategic partnerwith marketing, not just as a platform provider.

Be clear in what data you will need. Show the IT team what you willuse and what impact it could have. Be prepared to parse out what youneed versus what you want just for the heck of it. You’d be amazed athow much the IT folks can achieve once they are “in the know” andyou acknowledge how valuable their contribution is!

Ask for self-service access to the data. Too often IT builds a customer datawarehouse that can only be accessed through SQL queries. This is badfor IT because now each time you have a question about customers, orneed a segment, IT will have to do work. This is also bad for marketing,because with each request you submit you need to “wait in line” behindother projects.

Make sure IT knows your existing technology stack. If you ask IT for helpwith data integration, make sure you outline what you ultimately wantto do with the data. If you ask for data integration only, you may geta customer data warehouse that is hard to access (see #3) but also thatdoesn’t talk to your existing campaign execution tools.

Discuss ongoing requirements. Make sure that IT doesn’t only focus onthe initial development and deployment, but understands your needs forongoing updates. Customer data changes so quickly that, at a minimum,you will need daily updated customer profiles and segments. Data inte-gration and cleansing really is an ongoing process, and scheduling ITresources on a one-time basis is not sufficient. Too often, IT will delivera solution with too many manual steps, making it cost prohibitiveto update profiles and segments more than once a quarter. Insteadmarketing needs near real-time, and therefore automated, updates.

Start small and iterate quickly. Together you can come up with aroadmap that makes sense. Perhaps don’t ask for all data to be integratedat once. Start by capturing all digital data from email and your websitefor example, and get your feet wet with campaigns that use just thesedata sources. After you have proven success, it will be easier to justifythe incremental investment to add other data sources such as your storetransaction systems.


Be empathetic to the IT team. Acknowledge the challenges of the ITteam. If the IT team is struggling to put the data together, suggest asimpler ask. Have an idea of what your true deadlines are. Don’t be afraidto get a drink together. Reach out and ask for their ideas on what theythink you are missing, how you could simplify your marketing report ordata! Trust helps.

Get outside help. Introduce IT to some of your marketing vendors.They may have expertise in ongoing data integration, cleansing and anal-ysis. Data integration is a specialized skill and it may not be somethingthat your IT group does every day. Also, your IT organization may notbe familiar with these vendors so you could champion a solution.

Involve IT early. Early in the project, schedule a data discovery callwhere you involve the IT organization to define the use cases/businessrequirements for the data integration.

Assign dedicated resources. Treat data integration as a separate projectdeserving of its own project manager. The project manager can be inIT or marketing. In either case, make sure there is one specific, namedperson in IT to work with on scoping your project, selecting a vendor,and seeing the project to completion.

One Hundred Questions to Ask Your Data

Once you have all your data in one place, you can start to understandyour business and your customers better using this data. The list thatfollows is a place for you to start. The questions to ask are endless, butwe thought it would be useful to get an idea of some of the things youcan learn from centralized customer data.

Sales

1. How many new customers am I acquiring each month?2. What is our true cost to acquire new customers?3. What is my revenue per customer? How is it trending?4. How seasonal are my revenue and margin?5. Is most of my revenue coming from new or repeat buyers?6. Is most of my margin coming from new or repeat buyers?


7. What is my annual total number of orders of products I shipped?8. What is my order value by month: do some months see larger deals?9. What is my annual average order value and how is it trending over

time?10. How does my revenue break down by access device (mobile, tablet,

etc.)?11. How does my revenue break down by geography?12. How does my revenue break down by store or by sales representa-

tive?13. How much of my revenue come from non-marketable customers?

Customers

Product clusters

14. How many customers are in each product-based cluster?15. How much is each product-based cluster member worth?16. Which product-based cluster produces the most revenue?17. Which brand-based cluster produces the most margin?18. What channel does each product-based cluster prefer?

Brand clusters19. How many customers are in each brand-based cluster?20. How much is each brand-based cluster member worth?21. Which brand-based cluster produces the most revenue?22. Which brand-based cluster produces the most margin?23. What channel does each brand-based cluster prefer?

Behavioral clusters24. How many customers are in each behavioral cluster?25. How much is each behavioral cluster member worth?26. Which behavioral cluster produces the most revenue?27. Which behavioral cluster produces the most margin?28. What channel does each behavioral cluster prefer?29. What percentage of my customers are discount buyers?30. What percentage of my customers are frequent buyers?31. What percentage of my customers are full-price (high-margin)

buyers?32. What percentage of my customers are one-time buyers?


33. Who are my high-return complainers?34. Who are my seasonal customers?35. Who are my single channel customers?

Lifetime value36. Who are my most valuable customers?37. What is the (predicted) lifetime value of my top 10% of customers?38. What % of revenues comes from my top 10% (or bottom 10%) of

customers?39. What is the order frequency of my top 10% (or bottom 10%) of

customers?40. What brands do the highest spenders prefer?41. What product categories do the highest spenders prefer?42. What channels do the highest spenders prefer?43. How to define a VIP?44. How many high-value customers do I have that are at risk of

leaving?45. What is my share of wallet for each customer (by customer seg-

ment)?46. What is my upside in each and every customer?47. Which accounts have a high potential lifetime value, but a low

penetration?48. What is the predicted lifetime value by gender?49. Is the predicted lifetime value of bargain hunters lower?50. Is the lifetime value of mobile shoppers higher or lower?51. What brand preferences do my most valuable customers have?52. Is the lifetime value of loyalty program participants higher than

average?

Likelihood to buy53. What is the revenue (and margin) impact of offering free shipping?54. Did discounts drive incremental sales?55. Did discounts drive incremental margin—considering the costs of

the promotion?56. What are the best incentives to give to each of our customers?57. Which are my high potential leads of people coming in?58. Should I charge a membership fee?59. Which existing customers are most likely to buy again?


Life cycle stage60. How many active customers do I have (who bought in the past

12 months)?61. How many of my customers have lapsed?62. Is most of my revenue and margin coming from new or repeat

customers?63. How quickly will a buyer typically make their second purchase?64. How many customers can I re-engage with a replenishment

campaign?65. How many customers can I re-engage with a new customer

welcome campaign?66. How many buyers with known emails have been to my website

recently but did not purchase?67. How many customers have recently opened an email but did not

buy in a long time?68. Should I focus on retention?69. Are my new customers returning (and is this getting better or

worse)?70. How many multi-time buyers do I have and when did they last

purchase?71. Which customers are at risk of churn?

Demographics

72. How many individual households buy from me?73. What is the average order value by gender?74. What is the distance to my closest store for each customer

(segment)?

Recommendations75. What product to recommend to each customer next?76. What is the primary channel for each of my customers?

Marketing/Channels

77. How many shopping carts are being abandoned each month?78. How many web searches are being abandoned each month?79. What is the revenue per email and how is this changing over time?80. How many engaged subscribers actually read my emails?


81. Is my number of engaged subscribers growing or declining?82. What is the performance of my direct marketing campaigns?83. What programs are giving me the best return?84. Which promotions drive the most sales?85. Does my catalogue add profits?86. Which customers have not received any emails from me in the past

year?87. What percentages of my sales come from which channel?88. What percentage of margin comes from which channel?89. How are my sales trending by channel (annual revenue by channel)?90. How are my sales trending by channel (monthly revenue by

channel)?91. How is my margin trending by channel?92. Which channel gets us the most profitable customers?93. Which channel gets us the most loyal customers?

Products

94. What is my revenue and margin by product category?95. How many people can I target with my product introduction

campaign?96. What product categories are performing the best?97. What is the purchase frequency of certain product categories?98. Are people in a specific zip code area buying specific products?99. Which customers will be interested in this new product/

content/event, etc.?100. How many different products types does each customer buy

from us?

CHAPTER 4

Managing YourCustomers as a Portfolioto Improve YourValuation

The best way for any business to maximize enterprise or shareholdervalue is to maximize the customer lifetime value, or profitability, for

each and every customer. Customers are the most important assets for afirm and therefore customer lifetime value is the most important metricin marketing. If you maximize the lifetime value, or profitability, of eachand every customer, you also maximize the profitability and valuation ofyour company as a whole.

The best way to optimize lifetime value for any customer is to givethat customer the best possible experience throughout the life cycle oftheir interaction with your brand: from their first exposure to your brandto becoming paying customers, coming back a second time, and even-tually turning into loyal brand advocates for your business. Predictivemarketing explains how you can optimize the lifetime value for everyprospect and customer. Now that marketers have more detailed infor-mation on past, current, and future needs of customers at their fingertips,delighting each customer one at a time is finally possible.

63


The best way to optimize lifetime value for all customers is to adopt aportfolio approach. Marketers need to recognize that different groups ofcustomers have different value and different behaviors and take differentactions based on these distinct customer segments. We cover types ofclustering in another chapter, but changing the mentality from singlevalue focus to managing a portfolio of values is important. You needa different strategy for those customers who are at risk of leaving youthan for those who already have a high likelihood to buy. You will needto adjust your thinking and budget for customers that are unprofitablecompared with more profitable customers.

We sometimes hear marketers get the advice to “fire” unprofitablecustomers, or just focus on high-value customers, but this is a mistake.Every customer contributes to revenue, and when you acquire customersand manage a portfolio, there is always a mix. The important focus is themix and how this mix is trending.

What Is Customer Lifetime Value?

In general, customer lifetime value is a term that describes how muchrevenue or profit you can expect from customers over their lifetime doingbusiness with you. There are a couple of different ways to calculate anduse lifetime value, depending on the marketing problem at hand.

Historical Lifetime Value

Historical lifetime value or LTV for short is defined as the actual profits—gross margin minus direct costs—from customers over their lifetimeso far, adjusted by subtracting the acquisition cost of those customers.Note that historical lifetime value only takes into account past pur-chases, not future purchases. The only time to use historical lifetime,rather than predicted customer value, is when you are trying to detect ifthe customer value of a specific customer or customer segment is trend-ing up or down. A customer might have spent $500 two years ago, butonly $200 in the last year. It is this change in lifetime value that sig-nals underlying trends, risks, and opportunities. If a customer’s historicallifetime value is trending down, this is called value migration, and thiscan be an early warning signal of customers unsubscribing from yourservice or planning to stop buying from your website. Detecting value

Managing Your Customers as a Portfolio to Improve Your Valuation 65

migration allows you to catch customers before they walk out the doorand it is too late to win them back. Identifying a change in histori-cal lifetime value allows you to implement a reactivation or proactivechurn campaign to turn the tide.

Beyond value migration, customers may be changing their spend-ing habits in other important ways. A certain customer may have madeonly one big purchase last year, but this year they are making smallerpurchases more often. While the customer lifetime value of this per-son has not changed, your marketing approach and goals for the personshould change.

In order to accurately calculate historical lifetime value, you needto be able to link all purchases made by the same person—even if thatperson used slightly different emails, names, or addresses. The averageAmerican has three email addresses. If you are like most companies,you may have separate order databases for different channels. The ordersfrom the web are often recorded separately from the brick-and-mortarstore purchases and those are separate from sales made through the callcenter. Unless you can link purchases from these separate channels to thesame physical person, you will not have an accurate picture of lifetimevalue. For some products, it might be important to also understand thetotal value of a household or an account. Perhaps I spend only very littlewith a brand myself, but by acquiring me, the brand has also acquiredthe revenues from my spouse and children. When comparing the coststo acquire me to the revenues of my entire household, I could be a veryprofitable investment. The only way to do this is to make sure you canassociate family members of a household to each other. Similarly, in busi-ness marketing you need to be able to associate different buyers in thesame company with the master account to understand the true value ofa customer.

We recommend that you take the cost to service a customer intoaccount whenever you can in order to calculate historical lifetime value.This includes returns and discounts, as well as product cost. On average,9 percent of all retail sales in the United States are returned by consumers,so ignoring returns would skew the results. Some practitioners calculateLTV without the acquisition cost. If LTV is being utilized to make acqui-sition decisions, acquisition cost should be taken into account. However,if it is for existing customers, acquisition cost is a sunk cost and shouldnot be used.


Predicted Customer Value

Predicted customer value is the projected value, revenues, and costs,adjusted for the time-value of money, of a customer looking forwardseveral years. Your average retention rate will tell you how many years inthe future, on average, you will retain a customer and how many yearsof future revenues to take into account. We typically look one to threeyears ahead when calculating predicted customer value.

Predicted customer value is very useful, especially when decidinghow much money to invest in acquiring or retaining a specific customer.If you were to only look at historical lifetime value, you would signifi-cantly underestimate the potential of a customer and likely underinvestin the acquisition or retention of certain customers. It can also be usedto identify high-value customers very early in the life cycle. After herfirst purchase, a future high-value customer looks just like everybodyelse. If you could recognize the high-potential customer early in the lifecycle, you could start differentiated treatment right then and there andincrease the odds this high-value customer will stick with you.

One person might have just bought this expensive jacket but hemight have only been a customer for two months, but another customermight have been a customer for five years and bought the same jacket.If you were to look at historical lifetime value, you might draw the con-clusion that one customer is more valuable than the other. However,these two might very well become equal value customers and shouldprobably be treated in much the same way. If you look at historicallifetime value you look too much at old customers and will miss theopportunity to acquire or retain more recent, high-potential, customers.

With predictive analytics, you can estimate the future value of a cus-tomer by comparing a customer to the thousands or millions of othersthat have come before them. You can predict future lifetime value byfinding customers that look just like them. From the example we usedearlier, buying a certain type of jacket may very well be an early indicatorof a well-known pattern of behavior for a high-value customer. Even ifpredicted customer value is not accurate in absolute dollar terms, therank order it provides gives the marketer focus on the right segmentand trends.


Here are some examples of factors that can signal future lifetimevalue. Predictive marketing software typically looks at hundreds of factorslike these but will only use those that actually correlate with futurelifetime value in your particular company or situation:

• Recency of engagement: The recency of purchases, web visits,reviews, and email clicks may all be important predictors of futurepurchases and thus future customer value.

• Size of the first order: Customers who make a large first order aremore likely to end up being valuable shoppers.

• Discount on the first order: Customers who buy full price are morelikely to become valuable over their lifetime.

• Multiple types of products in the first order: Buying from differentcategories, such as shoes and electronics, in your first order is a signalof future customer value.

• Time between orders: Most valuable shoppers make frequent pur-chases and thus a shopper who places a second order quickly is morelikely to become a high-value customer.

• Time spent on website: The more time prospects or customers spendon your website, the higher their likelihood to buy and the highertheir predicted customer value.

• Social and email engagement: Customer engagement of any kind,including email opens and clicks or social engagement, are great pre-dictors for likelihood to buy and predicted customer value. Often itis not the amount of engagement that matters, but the consistency orfrequency of engagement. Spending a little time every day is a morereliable indicator than spending hours sporadically.

• Acquisition source: It turns out that certain channels drive highervalue customers than others. The customer who came from a fashionblog may have a higher predictive value than the customer acquiredthrough a banner ad.

• Geography: Customers in certain zip codes have a greater predictedcustomer value than others. Rural populations tend to be more stable,move less frequently, and therefore have more loyal purchase behavior.Zip codes can sometimes predict what type of products people buy.


For example, zip codes with many apartment buildings have a lowpredicted customer value for certain products, such as lawn mowers.

• Seasonality: Retail customers who are acquired during the holidaystend to be about 14 percent less valuable than those acquired duringother times of the year.

• Personal referrals: People who came to your brand through a per-sonal referral tend to be more loyal than those who buy because of anadvertisement.

Predictions about lifetime value are not destiny. Marketers can domuch to change the course of history here. Take, for example, thefact that shoppers acquired during the holidays tend to be less valuableand less loyal than shoppers acquired at other times during the year.One skincare company decided to focus its retention efforts on thisholiday cohort specifically. It set up an email marketing campaign toincrease brand loyalty among new Cyber Monday customers, sendingregular reminders for refills and recommending other products ofinterest. They were able to reverse the trend, and lifetime value of thesenew holiday customers is now 5 percent higher than the companyaverage. By focusing on specific outreach to underserved customersegments, the company was able to offer personalized promotions thatultimately drove greater brand loyalty. The important lesson is that oncecustomers are acquired, the best strategy is to focus on engaging themto grow and retain them, ignoring the cost of acquisition.

Upside Lifetime Value

Upside lifetime value, which is also called size of wallet, calculates howmuch more money a customer still has to spend with you. This is moneythat the customer is already spending at your competition to buy theproducts you offer. Algorithms can figure out size of wallet by com-paring a customer to other like-minded customers. It is important formarketers to focus on what size the wallet is, because it is always easierto grow a relationship with an existing customer than to acquire a newone. Unfortunately, most marketers have been taught to focus more onnew customer acquisition than on engaging and retaining existing cus-tomers. Especially if the customers have high upside potential, marketersshould focus on how to deepen their relationship by introducing them


to new products or serving them in a differentiated way. Very few com-panies calculate and utilize the upside or share of wallet potential of acustomer, yet it can be a very powerful way to identify customers tofocus on.

The critical difference between future lifetime value and share ofwallet is often in the types of products that are factored into the analysis.For future lifetime value, you tend to look at just those products that acustomer is already buying from you. For example, I may be a hockeyplayer, and I may be buying my hockey tape from a specific outlet everycouple of months. Based on this, the company can project that if theyretain me, I will buy a lot more hockey tape in the future and perhapshave a predicted lifetime value of $300. However, because I am buyinghockey tape, I am likely in the market for skate sharpening, hockey sticks,and occasional gear upgrades. I am clearly buying those things elsewhereright now. If I were to buy all of my gear at the same place I buy mytape, my future lifetime value is probably well over $1,000.

Let’s look at another example, this time in business marketing.An electronics company has quite a few customers who only buy inkjetcartridges for printers. They buy these cartridges regularly and spend$20,000 a year on average, leading you to think it’s a great customer.But the fact they are buying these high-end cartridges means that theyprobably also have a big office with servers, laptops, and other productsthat could use services or add-on products that you are not selling tothem. The fact that you are not selling those other products to them isa missed opportunity.

You can use share of wallet analysis to find upsell targets, as onebusiness software company did. It took all of its business customers andbroke out those that were similar in size and industry. Out of 100,000customers, they found 20,000 businesses in the insurance industry with100 to 150 employees. They then divided these customers into valuesegments. The top 25 percent of these small insurance companies spent$30,000 a year, the next 25 percent spent $10,000 a year, the third 25 per-cent spent $5,000 a year, and the bottom 25 percent spent $1,000 a year.All of these businesses are similar, and they all have a similar spend-ing potential. Maybe not all of them will be large customers spending$30,000 a year, but all should at least be able to spend $10,000 a year,or as much as the second group. This means that for all the customersin the $1,000 bucket, you have an upside potential to sell more products


and services of $9,000, and for the customers in the $5,000 bucket youhave an upside potential of $5,000.

Increase Customer Lifetime Valuefor One Customer

The customer life cycle is a term used to describe the progression ofsteps a customer goes through when considering, purchasing, using, andmaintaining loyalty to a product or a service. The customer life cycleemphasizes the individual journey of each customer and encourages mar-keters to think about the right approach to every customer. The life cyclemodel underscores the importance of customer repeat engagement witha brand. A marketer’s job is try to give customers the best possible expe-rience at any point during their life cycle or journey with your brandand—in doing so—increase the loyalty and the value of those customers.

Figure 4.1 gives an overview of some of the strategies a marketercould use in the acquisition, growth and retention phase of a customer’slife cycle. Some of these programs have significantly higher return oninvestment than traditional broadcast marketing techniques. Whereas theaverage open rate of a broadcast email is about 14% and the averagerevenue per email sent $0.05, the various campaigns that are triggeredby a customer’s life cycle can see open rates that are two or three timeshigher and revenue per email up to $6 per email, a whopping 130 timesthe average of a broadcast email.

Acquire

It often takes many interactions to acquire a new customer. Once youhave acquired this customer, you have engaged in a transaction, but youcan’t really speak of a customer relationship just yet. In some industries,like retail, 70 percent of customers never come back to buy a second

Acquisition Growth Retention

Look-alike targeting Repeat purchase program Loyalty appreciationRemarketing Cluster-based targeting Customer reactivation

Figure 4.1 Sample Life Cycle Marketing Strategies


time. Your primary goal when it comes to acquiring new customersit to acquire the right customers. Some people simply have a higherlikelihood to become large, loyal customers. When it comes to acqui-sition you can focus on acquiring these more loyal customers usinglook-alike targeting. You can also increase conversion of prospects tobuyers using remarketing techniques. Both will be discussed in greaterdetail in Chapter 11.

Grow

Your primary goal for newly acquired customers is to turn them fromone-time buyers into repeat customers. Once a customer comes backand buys a second time, you have now started a relationship. It changesthe dynamics completely. Whereas retention rates for one-time buyers inretail are around 30 percent, the retention rate for two-time buyers jumpsto 70 percent. A repeat purchase program and cluster-based targetingare just two examples of strategies that you might use to grow customervalue. Chapter 12 is entirely dedicated to customer growth strategies.

Retain

Even when you have maximized your share of wallet with a customer,your job is not yet done. Your primary strategy for loyal customers is torecognize and appreciate them. With social media and the Internet, thepower of word of mouth is stronger than ever. A loyal brand advocatecan bring yet more customer revenue by referring you to their friendsor family, or by raving about you publicly and influencing strangers toalso buy from you. Not all churn is created equal. It is especially impor-tant to retain your high value customers. Loyalty appreciation and lapsedcustomer reactivation are just two of many strategies you can use toproactively retain your customers. Chapter 13 will introduce variousretention metrics and methodologies.

In France, the telecommunication company Orange developeda comprehensive strategy to preserve its customer base, when a newentrant disrupted the market by introducing plans 50 percentage cheaperthan the existing ones. The marketing strategy included actions at everystage of the customer life cycle:


From an acquisition standpoint, the operator created new, moreattractive plans to adapt to its new competitor. On the one hand, thecompany built a new low price brand to target the digital and price sen-sitive segment (specific target of the new entrant). On the other hand,fierce competition of the other existing operators pushed the companyto reduce its standard plans by 20 percentage. The average revenueper user (ARPU), revenue and margin were at risk, since the existingcustomer had the opportunity to massively migrate to cheaper plans.

To limit value destruction, the firm developed a massive “Personal-ized Customer Program,” in which it contacted proactively customersevery 6 months with the promise to adapt their plan to their actual voice,text, and data consumption. Treatment was adapted to customer lifetimevalue: VIP customers were contacted by call centers whereas mediumvalue customers received an email and low-value customers a text mes-sage. Algorithms were used to develop personalized recommendationsdepending on the customer consumption, ARPU, and other factors.For instance, if a customer was paying more than the actual price ofthe plan due to international communications, the company recom-mended a more comprehensive (and expensive) plan including interna-tional communications. Cross-selling campaigns were also launched toequip mobile clients with Internet subscriptions in an attractive “quadru-ple play” bundle. This program resulted in a decrease in the ARPU ofonly 10 percent in a year.

Proactive retention was the most challenging program, since thecompany had a large customer base, hence a high risk of revenueattrition. The first step was to categorize the root causes for churn bycustomer segment and to deploy retention actions for each one of them.Models were developed to estimate the likelihood to churn by segment.The company analyzed the best time to reactivate a client: a customer wasconsidered at risk between the three months prior to the end date of hersubscription and the three following months. For price sensitive clients,top churn root causes were related to the price of the plan, in which casethey were offered to migrate to the low-price offers. Digital high-valueclients were leaving to buy the most recent smartphone at a significantlylower price when they churned to the competition (operators com-monly offer large discounts on mobile phones to new customers, inorder to increase their customer base). The company solved this issue byadjusting the smartphones’ prices for premium clients, while decreasing


the subsidy rates for lower value clients. Besides, it included a newfeature in its premium offers: one new smartphone for free every year.While it lost more than 700,000 customers in the first two quarters ofthe year, the operator preserved its customer base and reached more than800,000 net adds (acquisitions minus churns) in the two following quar-ters. (Source: www.orange.com/en/content/download/10703/237238/version/5/file/FY+2012+EN+VDEF.pdf; http://satisfait.orange.fr/maitrise_budget_bilan_conseil_personnalise.php.)

Increase Customer Lifetime Valuefor All Customers

Now let’s put all we have learned about customer lifetime value togetherin one framework. As a company you don’t just have one customer.To optimize enterprise value you need to optimize total customerequity—or the sum total of the lifetime value of all customers.

An easy way to think about optimizing customer equity is to thinkof your customer pool as a physical pool full of water. Think of increas-ing customer lifetime value across the customer portfolio as increasingthe water level in a pool. The pool is filled with the money spent by theactive customers of your brand. Active customers are those customerswho have spent money with you in the past 12 months. Higher valuecustomers spend more money and fill the pool up faster than lower valuecustomers. Water is draining as customers leave you and stop spend-ing money with you. Some customers are coming back after a hiatus.How do you increase the water level in your customer pool? You can:add more (valuable) customers, prevent churn (and value migration),or reactivate inactive customers. Figure 4.2 summarizes this methodol-ogy of optimizing customer equity across your portfolio using this poolcycle analogy.

Add More (Valuable) Customers

You need to add at least as many customers as you lose in a year, prefer-ably more. Therefore, how many customers you need to acquire in agiven year to at least stay stable depends on your churn rate and on yourrate of value migration. If you lose more customers, or if you lose moreof your valuable customers, you will need to acquire more revenues to

http://www.orange.com/en/content/download/10703/237238/version/5/file/FY+2012+EN+VDEF.pdf

http://satisfait.orange.fr/maitrise_budget_bilan_conseil_personnalise.php


Figure 4.2 The Pool Cycle Management Framework

compensate for this. The opposite is also true. If you can find a way toacquire more valuable customers, you don’t need to acquire as many ofthem. We talk about marketing spending optimization in Chapter 5 andgive examples of specific marketing strategies to acquire more valuablecustomers in Chapter 11.

Prevent Churn

When a customer leaves you it may be too late to get him or her back.Your chances of gaining a customer for life are much greater if you proac-tively reach out to a customer than if you wait for that customer to leaveyou and try to get them back later. We elaborate on strategies to growcustomer value in Chapter 12.

Engage Inactive Customers

When a customer leaves you, not all is lost. On average it is 10 timescheaper to reactivate a lapsed customer than it is to acquire a new one.Reactivation programs for lapsed customers therefore are a low-hangingfruit for marketers looking for new revenue streams. It is even better toengage a customer to prevent a customer from lapsing in the first place.We focus on programs to retain and reactivate customers in Chapter 13.

PART 2

Nine Easy Plays toGet Started withPredictive Marketing

75

CHAPTER 5

Play One: OptimizeYour MarketingSpending UsingCustomer Data

Optimizing marketing spend is a vast subject to cover on its own.However, most of the approaches lack customer behavior and pre-

dictions as their focal point, which leads to media response optimization,and not true marketing optimization.

When asked to allocate marketing budgets, most marketers imme-diately think about allocating budget to the most important sources ofrevenues, the best performing channels and the best performing key-words from a response perspective, ignoring the customer who respondsto marketing. However, we want to introduce a different framework tothink about marketing spending: the predictive marketing way to allo-cate spending is based on allocating dollars to the right people, ratherthan to the right products or channels.

As we have said in Chapter 4, your most important job as a marketeris to optimize the lifetime value of every customer, and the customerequity of your customer base as a whole. Therefore, you need to allocateyour budget by taking customer into account. In this chapter we look atthe following framework when it comes to allocating marketing budget:

77

78 Nine Easy Plays to Get Started with Predictive Marketing

• Create separate plans to invest in acquisition, retention, and reactiva-tion.

• Differentiate your spending on high-, medium-, and low-valuecustomers.

• Find the products that bring in the highest lifetime value customers.• Find the channels that bring in the highest lifetime value customers.

Invest in Acquisition, Retention,and Reactivation

To an extent, it is cheaper to reactivate lapsed customers than to acquirenew customers. It is even cheaper to retain customers than to reactivatecustomers. On average, according to AgilOne data, existing customersspend 83 percentage more and visit 60 percentage more often. It istherefore critical to be deliberate about retaining customers, such as con-verting one-time buyers as quickly as possible to repeat buyers.

When thinking about allocating your marketing spending, thinkabout how to allocate budget to your acquisition, retention, and reac-tivation efforts separately. If you can, you may even want to allocatedifferent marketing personnel to each of these efforts. Some companiesare starting to organize around groups of customers and the objectiveto be achieved with each group: acquiring new customers, engagingexisting customers, and reactivating lapsed customers.

In our experience, most companies are too focused on acquiringnew customers when they could achieve growth more cost effectivelyby focusing more on retention and reactivation. Retaining existing cus-tomers requires more work but ultimately is more efficient from a costperspective. Predictive marketing is a key enabler, as we outlined in thisbook to engage existing customers.

Let’s look at the growth numbers of a fictitious company (seeFigure 5.1). This company is consistent in acquiring new customers:it brings in about 175,000 new customers every year. However, everyyear it is getting better at bringing back existing or past customers.It brought in business from 266,000 existing customers in 2014,a 44 percent increase from the 185,000 existing customers it booked in2011. The growth of this company is almost entirely fueled by existingcustomers. This is usually a very good thing, especially because in ourexperience reactivating existing customers is about 10 times cheaper than

Play One: Optimize Your Marketing Spending Using Customer Data 79

Figure 5.1 Growth from New and Existing Customers

acquiring new ones. So not only is this company growing its top linerevenue, it is also becoming more profitable.

For the next year, the company can now project growth of its existingcustomers (see Figure 5.2). If growth of existing customers continues,then it may have 305,000 customers buying in 2015. That means it onlyneeds to acquire 136,000 new customers to achieve the same revenuesnext year as it did this year. The numbers used in this example are typicalfor the retail industry, and if you are in that industry you can use theseto do your own planning:

• About 40 percent of new customers will come back the next year asrepeat customers.

• About 70 percent of repeat customers will come back the next yearas repeat customers.

• Ten percent of past, lapsed customers can be reactivated to return ascustomers in the next year.

For most companies, acquisition requires 7 to 10 marketing touchesand each touch is three to five times more expensive (due to lower


Figure 5.2 How Many New Customers Do You Need?

response rates) than for existing customers. Existing customers requireonly three to five touches. The resulting economics are 10-20 timeshigher costs for customer acquisition than for customer retention.

Perhaps the most important point here is that every percentageimprovement in retention has a positive compound effect on growth,while acquisition starts from zero year after year and is very costly.Continue to focus on improving retention by measuring and improvingretention over time, improving as compared to your peers, and bybreaking down retention into its components and improving andbenchmarking the piece parts. For example, retention by geography orby product categories might be very different and offer opportunitiesfor improvement.

If we go back to the company in our example and benchmark themagainst their peers, it may no longer surprise you that they are outper-forming their peers when it comes to existing customer growth rate (seeFigure 5.3). They are, however, lagging slightly when it comes to newcustomer growth. Benchmarking against your peers is a powerful way toidentify areas of opportunity for your company.


Figure 5.3 Benchmarking Your Growth

When you engage in benchmarking, compare each of the customerlife cycle stages separately. For business marketing, this may be measuringthe conversion in each of the stages of the customer funnel. You mightfind that you are doing better than average when it comes to convertingprospects and suspects to leads for your company, but that you lag whenit comes to converting leads to trial or first purchase (see Figure 5.4).

When you decide how much to invest in each of your life cyclesegments, take into consideration that conversion rates for each groupare very different. When advertising to existing customers, you may findthat your conversion rate is 60 percent, whereas perhaps the conversionrate of advertising to brand-new customers is only 6 percent. This means

Figure 5.4 Engineering the Marketing Funnel


that you should be willing to spend more for a click from an existingcustomer than for a click from a new customer.

We give a detailed example in figures 5.5 and 5.6. In this example,let’s assume you are getting 100 visitors to your website using pay-to-clickadvertising. For every 100 visitors, you are getting 20 new orders. How-ever, of these 20 orders, 15 come from existing customers reactivated byyour campaign and only five orders come from brand-new customers.

This scenario lines up with what we see across businesses. Most ofyour traffic tends to come from new customers, whereas most of yourorders come from repeat customers. It may look at first like you aregetting a 20 percent conversion rate from your campaign, whereas inreality you are seeing a blended conversion rate from new and existingcustomers, each with very different economics.

If you had to spend $0.20 for each click, you would pay $20 for the100 visitors to your site or $1 per order. On the surface this campaign ismaking you money because you are making $2 in margin per order, so$1 of profit for each order. Because of the difference between new andrepeat customers, however, the truth is much different.

Figure 5.5 Marketing Funnel by Life Cycle


Clicks

Cost perClick($)

MarketingSpend

($) Orders

Cost perOrder

($)

AverageMargin

per Order($)

VCMargin

($)

Overall 100 $0.2 $20.00 20 $1.00 $2 $1.00New/

Prospects75 $0.2 $15.00 5 $3.00 $2 ($1.00)←Acquisition

CostExisting 25 $0.20 $5.00 15 $0.33 $2 $1.67 ←Net Profit

Figure 5.6 Marketing Spending Optimization Example

In reality, of the 20 orders received, 15 were from existing customers,which were reactivated with your marketing campaign, and only fivefrom brand-new customers. Looking at it this way, it turns out that thenew customer acquisition campaign is actually losing you money in theshort run. Each new customer costs you $3 to acquire and brings inonly $2 per order. This means you are losing $1 on the first transactionof each new customer. Losing money on acquiring new customers isnot necessarily a bad thing as long as you know what you are doing.Most marketers are fine losing money on customer acquisition, providedthat the customer has a lifetime value that will make the relationshippositive in the long run. We look at that next.

Each existing customer is costing you only $0.33 to reactivate and ismaking you $1.67 in net profit. That seems pretty good, but perhaps youcould have reactivated this customer with a well-timed email campaignand avoid the reactivation cost altogether.

The solution is to do optimization separately for your acquisitionand retention budget. You can optimize your acquisition budget basedon the time it takes to break even or the customer value achieved in thefirst 90 days. Optimize your retention budget based on net profit perorder, cutting down on negative net profit activities. For each of theseyou can use either actuals from the past or predicted measures for thebreakeven time period.

Optimizing Your Acquisition Budget

We recommend that you optimize your acquisition budget based on pay-back time or customer value (predicted or actual). Figure 5.7 illustratesthe concepts of payback time and customer value. The payback time isthe period it takes for you to recoup your acquisition spending. Let’s say it


Figure 5.7 The Customer Value Path

costs you $5 to acquire a customer and it takes you one month to receiveenough profit, or customer value, from that customer to recoup thatinvestment. The payback time for the customer acquired from this par-ticular activity is thus one month. You could rank all your acquisitionactivities based on payback time and pick the best ones.

The other way to optimize acquisition activities is to choose a timeperiod, for example 90 days, and to see which acquisition activitieslead to the greatest customer value at the end of the evaluation period.We picked a relatively short time period here, 90 days, because the acqui-sition landscape is very dynamic. Keywords that work one week may notwork the other week, so looking too far into the future may not leadto accurate results. In other words, the acquisition activity that leads togreat results this year may no longer exist one year later.


We recommend that for each of your acquisition efforts, you tryto fill in the payback time and the return on investment (ROI) fromcustomer value at the end of your evaluation period. You can use thetemplate in Figure 5.8 to do this. Then simply pick the best performingactivities for next year. Certain considerations go into play, such as thequantity of customers acquired, the variability over time, and the abilityto invest more.

Some acquisition sources, such as banner ads on a high-fashion blog,may have extremely short payback periods and yield a profit on the firstorder, but are contributing very few new customers. The marketers goalis not only to be optimal in the profit sense, but in quantity as well.Acquisition sources that might be highly efficient might yield only afew customers, and hence are not impactful. In one case, a blogger wasdelivering very valuable customers to a brand, but could only deliververy few customers with no ability to scale with a larger investment.

Second, some sources deliver consistent performance over time,whereas others have more variability. Consistent sources are alwayspreferred over highly variable ones. Lastly, you should look at yourability to invest more. Certain marketing investments such as searchengine marketing (SEM) can be increased if performing well, whereasothers might be supply constrained. Therefore, even though a source isperforming, there might not be a chance to invest more in these sources.

Optimize Your Retention Budget

We recommend that you optimize your retention budget based on con-tribution margin and customer value. After optimizing for acquisition,marketers need to build profitable relationships with those customersover time. After a customer has placed her first order, the acquisitioncost is already a “sunk cost.” Therefore, the approach should no longer

Source ofAcquisition

AcquisitionCost ($)

Payback Time(Days)

90-Day ROI(LTV-ACQ/ACQ)

SEM_Yahoo 12 43 15%Banner_FB 19 123 −20%Retargeting_SH 25 80 2%

Figure 5.8 Acquisition Sources Worksheet


take the acquisition cost into account, since those customers are alreadypaid for and optimized for that cost. This requires careful understandingof the variable contribution margin. The variable contribution marginis defined as the gross margin minus the marketing cost to generate thatorder. You should perform this calculation for each source of market-ing bringing orders from customers placing their second or subsequentorders. The contribution margin is going to be different for low-value,medium-value, and high-value customers. We look at allocating budgetto each of these groups of customers next.

Differentiate Spending Basedon Customer Value

Losing a high-value customer is much more costly than losing a low-value customer. Therefore, when allocating your retention budget thinkabout different value-based segments and how likely they are to buyfrom you again. For this you will need to be able to calculate thelikelihood to buy for each segment. The customers with a lower likeli-hood to buy are at greater risk of leaving you—and never make anotherpurchase again.

You should allocate budget based on customer value and risk.We will walk through the example of Figure 5.9. Let’s say you have600,000 new customers with an average annual value of $100. In addi-tion, your existing customers include 90,000 high-value customersthat spend an average $400 a year, 540,000 medium-value customersthat spend an average $110 a year, and 270,000 low-value customers that

Figure 5.9 Spending Based on Value and Risk


spend only $20 a year. Losing one of the $400 customers is more costlythan losing a $20 customer.

The historical retention rate for each of these groups may be differentas well. Perhaps your high-value, loyal customers have a 90 percent reten-tion rate as compared to a 20 percent retention rate for the low-value,discount shoppers. If your retention rate drops, you have $32 millionat risk with your high-value customers, $38 million at risk with yourmedium-value, and only $1 million at risk with your low-value cus-tomers. Another way to look at it is that if you can further increasethe retention rate of each group, you have $3.6 million of upside withhigh-value customers, $24 million of upside with medium-value cus-tomers, and $4.3 million of upside with low-value customers. So whereasthe retention rate of high-value customers is high at 90 percent, still37 percentage of all money at risk due to churn comes from this group.On the other hand, while the retention rate of your lowest value cus-tomers is only 20 percentage, only 2 percentage of all money at riskcomes from this group. It would make sense to spend more money toprotect the potential loss of medium- and high-value customers in thiscase. We share concrete tips and examples for campaigns to grow andretain customers in Chapters 12 and 13.

Find Products That Bring High-ValueCustomers

Now let’s look at your product portfolio through the lens of customers’lifetime value. If you simply rank your products based on the revenuethey bring in you may miss important insights. Let’s assume you sellfive different categories of products: category 1 through 5 as shown onFigure 5.10. Category 1 may bring in 25 percent of your new customerrevenues as compared to only 6 percent for category 5–type products.However, if you could only look two years into the future, it would beclear that it is in fact a different product category altogether, category2 that outperforms all other product types based on customer lifetimevalue. Some categories of products may have lower customer revenue inyear one but over time will result in more repeat purchases and highercustomer lifetime revenues. When deciding how much money to investin each product category, you should take future revenues into account.You can also spend more money to acquire category 2 customers becausethose have a higher lifetime value over time.


Figure 5.10 Lifetime Value by Acquisition Category

Find Channels That Bring High-ValueCustomers

Acquisition channels, too, can be ranked based on the types and valueof the customers it delivers. Not all channels deliver customers of equalvalue. In one example, a company uses various outlets to acquire newsubscribers. After 60 days, the customers acquired via the Spa channeland those acquired via DailyCandy or AdFusion are about eight timesmore valuable, on average, than those acquired via US Weekly.

Similarly, when Anne Swift, chief marketing nut at Nuts.com, anonline retailer of nuts and dried fruits, had to decide how much moneyto spend on specific adwords. She was somewhat reluctant to bid muchmoney for certain keywords. However, when she switched from lookingat just the first order to looking at predicted lifetime value to determineprofitability of a campaign, she found that many of her bidding strategieshad been too conservative.

When Adam Shaffer worked for PCM, a billion-dollar electronicsdistributor, he had to decide which sources to spend on. He initiallyfocused on adwords campaigns. However, he found that some of thesecampaigns were acquiring the same customers over and over again.Instead of spending one time to acquire a customer, customers were


searching and clicking paid links each time they came back for a repeatpurchase. These customers were not nearly as profitable as Adam hadinitially thought. On the other hand, when Adam looked at his directmail campaigns he was pleasantly surprised. Although these campaignswere very expensive, for certain segments of customers, direct mail wasa very profitable way to reach these high lifetime value customers.

In another example, we’re working with an online golf retailer toreview the performance of their Google Product Listing Ads (PLA).On the surface, it appeared that the Taylor Made and Adams PLAsgenerated the same number of buyers, with slightly more revenue andhigher margin customers coming from the Adams PLA: Taylor Made andAdams each brought 67 buyers, with $11,453.18 coming from Adamsand $10,855.01 coming from Taylor Made buyers. So on the surface theAdams program is the more successful one by a small margin. By look-ing at the same information by customer order count, we can see thatthe Taylor Made PLA acquired 20 new customers and drove 39 incre-mental orders from existing customers. The Adams PLA, however drove32 new customers and 26 incremental orders from existing customers.While Taylor Made drove fewer new acquisitions, it drove acquisitionof highly profitable customers: revenue per new customer was $678.72and margin $237.44 for each new Taylor Made buyer as compared to$326.48 in revenue and $174.29 per new customer for Adams. In con-clusion, the Taylor Made program probably is the more important oneto double down on. Existing customers can be targeted through otherchannels such as email, and while there were fewer new customers com-ing from the Taylor Made program, the overall new customer revenueand margin from this program outperformed the Adams program by asignificant margin.

The Case for Last-Touch Attribution

The biggest problem when finding channels that perform is that mul-tiple marketing efforts, in different channels, all assist in the outcome.The sheer number of available channels, tools, and keywords haveexponentially increased, leading to confusion and frustration frommarketers as to what marketing touch is responsible for what customeraction. For example, a customer clicked on a banner, three days latersearched and came to the website through a Google Adword ad, then


one day later, got an email and clicked. In such cases, all these marketingtouches contribute to the conversion, but in a fractional manner.Understanding this fractional contribution is the domain of revenueattribution.

In order to do this, marketers needed to attribute gross margin andcost of the specific activity with each marketing touch. It was quiteeasy when touches were fewer and last-touch attribution worked quitewell. With the advance of keyword advertising, retargeting, and variousother online campaigns, marketers started to demand more sophisticatedmarketing attribution, such as multitouch attribution. Nowadays, manyorders are touched by three-plus marketing events, which requires themarketer to sort out how much credit to give to each event.

Multitouch attribution can become complex very quickly, usingtechniques such as linear, time-decay, or time-window–based attribu-tion. Each of these methods can be the right solution. For example, linearattribution—where each marketing activity gets equal credit—could bea good way to give partial credit to customers clicking on advertise-ments, but if you are sending them emails, direct mail, that is, outboundtouches, the more you send the more credit they will get, hence reducingthe effectiveness of all together.

There certainly is value in multitouch attribution for consideredpurchase items, such as cars, home goods, insurance, where the buyingcycle is relatively long and the consumer is being educated and nurturedto make the right choice. For short sales cycle, impulse purchases, theproblem is less important.

Many marketers obsess over accuracy of such calculations. The rightapproach is to find the attribution approach that is as simple as possible,yet accurate enough to sort through good and bad marketing invest-ments. Test the following: How does the order of performance changebetween sources when you change attribution methods? If a market-ing source is performing poorly, does it all of a sudden perform wellwhen you switch from last-touch to linear attribution? Even if you do seechanges in rank order of sources, the next question to ask is the impor-tance of those sources in terms of the budget spent on them and theirperformance in variable contribution margin. If the order changes, butthose specific channels contribute little in the grand scheme of things,then maybe investing in complicated attribution modeling is not worthyour time after all.


In our experience, if you are marketing short decision-cycle productsand services, such as buying a pair of headphones or a laptop case, last-touch attribution will be perfectly accurate enough. Because the decisioncycle is short, consumers tend to make a decision whether to buy quickly,not requiring them to come back many times. Therefore, the last-touchattribution is more than adequate in providing the necessary accuracycompared to the marketing cost/benefit of using attribution and the costof owning and operating a sophisticated attribution system.

Often the difference between last-touch and multitouch attributiondoes not make a good source perform poorly, or a bad source performwell all of a sudden. Rather the okay performers, or sources that aremore or less zero, will flip-flop. We have tested the difference in rank-ing between last-touch and multitouch attribution and found only smalldifferences. In attributed revenues in dollar terms the differences wereless than 10 percent, and often much smaller given any specific channel.In terms of the rank order of the channels you can see the results of onesuch test in Figure 5.11. There were only two instances, both markedwith ∗ in Figure 5.11, where two similarly performing channels werereversed in order when switching attribution methods.

One reason that last-touch attribution works so well in many casesis that the time frame from marketing touch to customer decision hasshrunk from weeks to days, minutes, or even seconds. In the case of directmail, consumers typically respond with a purchase within weeks afterreceiving the catalog. Catalog marketers call this the matchback or back-match period. For most catalog campaigns, within four weeks 55 percentof all responses have come in. After eight weeks 85 percent of responses

Last touch attribution Multi-touch attribution

1. Email 1. Email2. Yahoo ads 2. Yahoo ads3. Adwords 3. Adwords

∗4. CJ ∗4. Direct mail∗5. Direct mail ∗5. CJ6. Performics 6. Performics7. Amazon 7. Amazon

∗8. FCBI ∗8. Skymall∗9. Skymall ∗9. FCBI

Figure 5.11 Last-Touch versus Multitouch Attribution


have come in and after 12 weeks 99 percent of responses have come in.Given this, most marketers would use four weeks as their matchbackwindow and only attribute responses that come in within the first fourweeks to the direct mail campaign.

Modern marketing campaigns have much shorter matchback win-dows than direct mail campaigns. The timescales in digital marketingtend to be days instead of weeks. This is why last-touch attributionworks well for digital campaigns because the customer action happensso quickly after the campaign. Using last-touch attribution can makefinding the channels that bring the best customers significantly easierand attainable for all marketers, even those who have not invested insophisticated marketing attribution solutions.

CHAPTER 6

Play Two: PredictCustomer Personasand Make MarketingRelevant Again

As we saw in Chapter 2, clustering is the automated, machine-learningpowered version of segmentation. Clustering is a powerful tool that

allows us to discover personas or communities within your customerbase. You segment customers to identify homogeneous groups thatexist within your customer base, which can be used to optimize anddifferentiate marketing actions or product strategy.

An online retailer we work with is operating an e-commerce storewhere triathletes, hobbyists, and cyclists can celebrate and support theirpassions. The company wanted to cater to individual customer interestsin a scalable and feasible way for its modest marketing team. It beganby looking at the different personas that make up its customers and dis-covered certain distinct communities around its products: professionalremote control racecar drivers; hobbyists building kits with their chil-dren; and remote control airplane enthusiasts. With this information,the company started to send out different newsletters to the differentpersonas. Addressing each of these groups with meaningful personal-ized content immediately paid off: email click rates increased 66 percent.

93


This is very typical, as we have seen similar increases in response withmany companies we work with.

Clustering is a means to an end and it is a tool to develop strategies.Strategy informs segmentation and not the other way around. Thethree example schemas we chose here are customer behavior aroundproducts, their shopping behavior, and their attitudes against variousbrands. Other schemas would depend on the business strategy you’retrying to develop. Just like we’re clustering customers, you could clusterkeywords, products, stores, employees, and any other entity in yourmarketing ecosystem. For example one of our customers, Arcelik, awhite goods and electronics manufacturer and retailer in Europe, wantedto improve store performance and understand the customer profile ofits stores. Head of CRM efforts Bora Cetiner with direction from theirVP of Marketing Tulin Karabuk clustered their stores and understoodwhat made them different. Equipped with this, they were able to makechanges to store layout, advertising, CRM efforts. Even the goals setfor each store depended on which cluster they belonged to.

When you utilize clustering for customers, we call the outcome clus-ters personas, because it captures the underlying persona for the customerthat belongs to that cluster. But how do you express personas in a waythat the marketer can understand the persona? The easiest way to dothis is using the same heuristics of how we humans do this. Humansare essentially difference engines that look for “edges” to detect a pic-ture or highs and lows that are different and keep only this information.For example when we look at a landscape, we notice the horizon, thesun setting, and we don’t pay attention to all the waves in the ocean.We use similar mechanism to display personas, and we express a cluster“DNA” to do this. Cluster DNA is not an industry standard term (yet),but we use it to describe the clusters to marketers. Cluster DNA depictshow a customer belonging to a specific cluster more or less likely prefersa product or behavior (or whatever the clustering schema is) versus anyother one.

Types of Clusters

Let’s look deeper into the three types of personas that we will useas our examples: Product-based clusters, brand-based clusters, andbehavior-based clusters. These are three examples that we utilize on afrequent basis and informs a wide variety of strategies.

Play Two: Predict Customer Personas and Make Marketing Relevant Again 95

Product-Based Clusters

Product-based clustering models group customers based on what typesor categories of products they tend to prefer and what types of productscustomers tend to buy together. Product-based cluster models are some-times also called category-based clusters. A product cluster can be broad orvery specific. In Figure 6.1 you can see people in one customer segmenttend to only buy sweaters, whereas those in another customer segmentbuy different types of active wear. These people might buy swimwearand watches, but never kids’ clothes, intimates, or jewelry. This is usefulinformation when deciding which product offers or email content tosend to each of these types of customers.

How could you use product clusters? A big sports retailer, which sellsathletic gear and apparel across different sports to a range of ages, initiallythought that the same female customers that were buying kids’ soccergear were also buying yoga clothes for themselves. The company begansending direct mail advertising yoga gear to the women who bought soc-cer products. However, when we ran our predictive clustering models,we quickly discovered there was no crossover between soccer moms andyoga moms at all. We found crossover shoppers in different categories ata much higher rate, which allowed the business to shift its focus awayfrom marketing to what they thought were soccer and yoga moms tothe newly discovered and much more qualified segment of soccer andbaseball moms.

Brand-Based Clusters

Brand-based clusters tell you what brands people are most likely tobuy. It groups customers together that prefer a group of brands morethan other. For instance, you will be able to identify which customersare likely to be interested when a specific brand releases new products.The models can also offer broader insights into related brands that maybe of interest to a customer by comparing her preference to an existingbrand cluster. The results of the model in Figure 6.2 illustrate thatcustomers who like Tahari by ASL also tend to like Calvin Klein andNine West, but would not be interested at all in Desigual or 6126.

Many retailers we work with have discovered that the brand affin-ity of customers tends to be stronger than their product-type affinity.This means that if a marketer in the example of Figure 6.2 sends apromotion of Nine West shoes to a customer who has bought a lot of


Figure 6.1 Example of Product-Based Clusters

Calvin Klein in the past, this promotion will likely generate a lot of sales.It will probably generate more sales than a campaign that is promotingwatches to people who belong to the active wear cluster and have boughta lot of sunglasses in the past. This is a generalization and it is importantto note that there are exceptions to every rule. This type of affinity is a


Figure 6.2 Example of Brand-Based Clusters

broader stroke affinity uncovering larger-scale patterns in your customerbase than product recommendations we will cover later.

Behavior-Based Clusters

A behavior-based clustering model groups customers based on how theywill behave while purchasing: Do they use the website or the call cen-ter? Are they discount addicts? How frequently do they buy? How muchdo they spend? How much time will pass before they purchase again?This algorithm helps set the right tone for contacting the customer.For instance, customers who only buy with heavy discounts may begreat targets for inventory-clearing sales, whereas customers who typi-cally pay full price would be better targets for a sneak-peek promotionof a new product line.

Behavior-based clustering can help identify completely new clustersyou didn’t know you had before. A clustering algorithm looks at a largenumber of variables including, but not limited to, average order size, daysbetween orders, first order revenue, order variety, discount sensitivity,order frequency, total number of items purchased, total number of ordersplaced, number of returns, products bought in the first order, and orderseasonality. Algorithms can look at hundreds of variables at the same timeand discover which variables and attributes actually matter. Typically aclustering algorithm will find six to eight statistically significant personasin your customer data.

Every behavior-based cluster will have a cluster DNA that will revealto the marketer which customer attributes are most differentiating andrelevant in order to group customers. Algorithms are able to segmentcustomers based on many more variables than a human being could. An


algorithm will start with hundreds of dimensions and finally may pick 20or so to define a certain behavioral cluster or persona. The example inFigure 6.3 shows some of the factors that could make up a cluster’s DNA.

Let’s look at an example from business marketing. Let’s assume youare selling tools to small homebuilders and contractors. Your behavior-based clusters could reveal that you have some customers who onlypurchase from your call center agents, but always after spending a con-siderable amount of time doing research online. You may have anothercluster with customers who stock up only once a year in your store,after receiving your direct mail postcard announcing your annual sale.And you have a third cluster, which also buy in your store, but basicallycome in once a week and always buy full price. Your marketing strategiesfor each of these three clusters would be very different. If you had onlyone customer behaving in this way, it would not be a cluster and notworth building a strategy around. However, the whole idea of clusters isthat there is a statistically significant group of customers who behave inthe same way.

Figure 6.3 Example of Behavior-Based Clusters


Airlines’ behavior-based clusters take into account factors like orderfrequency, days between orders, order seasonality, and discount sensitivityto help differentiate business travelers from leisure travelers. For airlines,this is a critical segmentation tool. Pricing and promotional strategies,as well as customer service strategies, can now be differentiated andtuned based on the types of customers being served on certain routesand flights.

Retailers might encounter behavior-based clusters such as full-priceinfrequent buyers, buyers who return products frequently (the so-called“returnaholics”), buyers with few orders who mainly buy when there isa large discount (the “discount junkies”), and buyers with a high orderfrequency who are very likely also your VIP customers. These differ-ent customer groups will be attracted to very different promotions.Your whales might enjoy receiving frequent emails from you and willpurchase almost every time. However, sending the same email and con-tent to your full-price infrequent buyers may turn them off to the pointthat they opt out from your email list. You might be better off sendingthis group a postcard if and when you are releasing a new product.

Using Clusters to ImproveCustomer Acquisition

Clusters are not only useful to target existing customers with more rel-evant communications. Clusters can also be used to craft more rele-vant new customer acquisition campaigns. Everything from the creativedesign to the places you advertise is influenced by the persona you arelooking to target and acquire. The better you understand the differentcustomer personas that buy from you, the more accurate you can be indesigning campaigns to attract more buyers just like it.

A leading vitamin and wellness brand used to collect data, like statis-tics on the best-selling products, and analyze it around the boardroomto make hypotheses about their business. For example, when a certainjoint supplement became a big seller, they assumed it was because of anuptick in their senior-aged customer base. They leveraged this hypothesisto tailor their campaign artwork, print mailers, and their media ads to anelderly customer. However, when they started to apply predictive clus-tering models to their data, clustering instead showed that the increase injoint supplements was actually due to a completely different customer:


customers that identified as body builders. Imagine how this impactedtheir business decisions! To start, the brand’s marketing agency madea swift change to the artwork and media plan to save the resources thatwould have been wasted marketing to an elderly customer. The companycan now work with its digital agency to make better product recommen-dations on e-commerce, as well as develop lifestyle content around thisaudience that is highly relevant to them for the blog and social channels.The brand’s store planners can merchandise the joint supplement along-side the protein powder in-store. Its PR firm can plan a highly targetedevent and bring in the right influencers.

Clusters can also be used successfully in combination with lookaliketargeting on social media and in display advertising. We will discuss thisuse case in more detail in Chapter 11.

Things to Watch Out for When Using Clusters

The biggest mistake that marketers make when it comes to segmenta-tion is to use only one-dimensional segmentation. No person belongsto only one segment. Segmentation is very contextual. Depending onthe situation, a customer belongs to a different segment. For example,from a product perspective, John could be a runner and Mary couldbe a swimmer, but John could also be a discount-sensitive buyer alwayshunting for deals from a behavioral perspective and Mary always buys thelatest product when it first comes out at full price. Now these are twodimensions that we can segment on that cannot be combined in a singledimensional way because another customer, Jane, could be a swimmerand discount-sensitive as well.

It is also important to find a way to update and use clusters on anongoing basis. Many companies spend a lot of time creating the perfectclustering algorithm, performing exhaustive analysis whereby to selectthe cluster that would bring the most value or action-ability. Buildinga cluster this way can easily cost $100,000 or more. What’s worse, bythe time the analysis is complete and in the hands of marketing, theinformation in the customer file is ancient history. Paying continuallyfor re-scoring a customer file on a quarterly (monthly usually is costprohibitive) basis is expensive and still leads to sub-optimal results. Thesedays, there is a better way, and modern software can ensure that clustersare updated and ready to use on a daily basis.


Clusters in Action

A global manufacturer and distributor of high-quality vitamins andnutritional supplements works to enhance the well-being of its cus-tomers by delivering high-quality, best-value nutritional supplementsand wellness products. The retailer’s European operations, which havemore than 1,000 stores, wanted to gain a deep understanding of bothonline and brick-and-mortar store transactions. In particular, the retailerwanted to learn which products customers buy, how frequently they buy,and how discounts motivate them to buy. They focused on what buyerbehaviors are driving the business, tailoring email campaigns accordingly.The retailer learned that the most valuable shoppers, which account foronly 20 percent of its customer base, drove more than 80 percent of itsprofit. It also learned that food drives customer frequency and attractsthese high-value shoppers. Certain food lines were considered unprof-itable before understanding this insight. Now the retailer continuesto carry these products to encourage customer frequency and attracthigh-value shoppers.

The company also learned that the purchase frequency of most of itscustomers is very low. However, loyal customers will shop more oftenand spend about 30 percent more per transaction. Also, customers whobuy the store magazine spend 50 percent more than those who don’t getthe magazine.

The insights helped the company make adjustments to its merchan-dising and marketing plans, increasing overall revenue by 1.5 percentand the revenue of a specific new product announcements by 4 percent.It was also able to record a rise in the number of customers, the aver-age order value, the average transaction value, and number of units pershopping cart.

The most innovative and impactful use of clustering is always start-ing with the end in mind. Define and understand the problem you’retrying to solve or strategy you’re trying to develop, then find the mostappropriate clustering schema that will help you tease apart personas towork with.

CHAPTER 7

Play Three: Predict theCustomer Journey forLife Cycle Marketing

In Chapter 4 we recommended that you look to optimize the lifetimevalue of each customer. In this chapter we look at the life cycle of a

single customer in more detail to see how our strategy should includeall stages of customer life cycle from acquisition to growth to retention.We start by looking at a simple customer life cycle or customer journeyand then see how segmenting customers based on their life cycle stagecan yield great insights and opportunities for customer growth. We alsogive you an overview of life cycle marketing strategies.

The Customer Value Journey

The basic principle of optimizing customer lifetime value is the samefor all stages of the life cycle and can be summarized in three words:give to get. Customers are much more likely to buy from you if theyhave a relationship with you. The best way to develop that relationshipis to deliver an experience of value. So to get customer (lifetime) value,give customer value.

Figure 7.1 shows basic concepts of the customer value journey.Customers will disengage you if they are not getting value from therelationship. The definition of value is very different from company

103


Figure 7.1 The Customer Value Journey

to company and it ranges from monetary value, especially in businessmarketing, to utilitarian value such as when using a great vacuumcleaner, to the self-confidence you may feel when wearing a designerdress. Whatever your definition of value, if customers are not getting it,they will look elsewhere.

There are three basic forces at work to create a relationship with acustomer. First, does it make sense for the business to have a relationshipwith the customer, which requires thinking through size and type oftransaction, future revenue streams, cross-sell, upsell opportunities, andcost of acquisition? Basically, will this customer deliver enough economicvalue to the business over time? Second, you need to think throughwhether it makes sense for the customer to have a relationship with you;this includes emotional attachment, risk of decision, follow-on products,

Play Three: Predict the Customer Journey for Life Cycle Marketing 105

or services required, involvement with the product, and so on. Basically,does our business offer what the customer is seeking? Third, to develop arelationship with a customer requires the business conditions to be rightfor the life cycle marketing to work. You need to be able to configurecustomer experience at a granular level and have the data to understandand predict customer behavior.

It is easier for companies selling certain products to develop a rela-tionship with its customers than for others. For example, it is really hardfor a deodorant brand to enable a life cycle relationship with a customer.Typically deodorant simply is not a high-engagement, high-risk productthat makes sense for the consumer to provide information about and getsomething in return. On the other hand, for a camera brand or retailer,building a relationship with its customers is a lot easier. The camera buyerwill engage with the brand starting from a camera purchase, continue tostay engaged with the brand for additional accessories, training videos,workshops, repairs, and so on. In this example, it really makes sense forboth the business and the consumer to engage in a relationship, whereit potentially benefits both and the business conditions are ripe to do so.

There are exceptions to every rule. For example, Secret is anantiperspirant/deodorant for women manufactured by Procter &Gamble (P&G) and first launched in 1956. In 2010 P&G launchedthe anti-bullying campaign “Mean Stinks” around the Secret brand.This campaign, now in its fifth year, has consistently drawn high engage-ment on social media and managed to really engage girls with the brandin a meaningful way.

First Value

As Confucius said: “Every journey starts with the first step,” and this istrue for the customer value journey as well. The first step in every cus-tomer journey is to get your prospective buyers to first value. We some-times call this the “first wow.” First value may occur after somebody hasbought from you, but it is even better to try and get your prospects tofirst value before they have paid you a penny. This is truly “give to get”marketing.

Consider the many ways in which you can deliver value beforesomebody is technically a customer. In consumer marketing, you maybe able to expose consumers to your brand via a great advertising


campaign, viral marketing campaign, or other piece of content. Perhapsprospects first learn about your clothing line from your fashion blog orattend an enjoyable fashion show. Many yoga enthusiasts and runnersfirst learn about active wear retailer Lululemon from the yoga classesand runs it offers at its stores. Marketers of time-share apartments aremasters of designing first value experiences in the form of free events atexotic locations.

Dominique’s daughter was recruited to a competitive soccer teamusing the “give to get” approach. While her parents had no intentionof signing her up for such a time- and money-consuming activity, theclub was smart and invited her daughter to participate in a free soccerclinic. From there she was invited to join the training of the team for acouple of weeks. Before long, she had made some great friends and wasbegging her mother to join the team. This is more than five years ago,and she is still playing (and her mom still paying).

Consumer goods companies have been using samples for a longtime to get you to experience their product before buying it. The foodsampling at warehouse club Costco is a popular pastime for many fami-lies. Retail companies are increasingly considering themselves content orentertainment companies first and retail companies second. Some retailCEOs even refer to themselves as editors-in-chief. The content is free,but the clothes are not.

Other brands are taking “give to get” very literally and engage incause marketing. The Burlington Coat Factory partners with K.I.D.S/Fashion Delivers, an organization that donates clothing and other prod-ucts to underprivileged families across America. Every store opening iscelebrated with a coat drive, making a charity event the first experiencea prospective shopper has with Burlington Coat Factory.

There are plenty of examples in business marketing as well. When wewere contemplating hiring a consultant, he first invited us to a free sem-inar, then offered a free one-hour consultation. He then sold us to afull group class and we eventually signed up for his services. Many busi-ness and consumer software companies now offer a free trial period or afreemium version of their software. Both allow prospects to use the fullyor partially featured product for a period of time before committing to apurchase. When offering a free trial period, make sure customers reachfirst value before the end of the trial period. Again, the definition of first


value is different for each company. If you are selling quality assurancetest software, perhaps first value is defining your first test that runs withat least two different device types. If you sell marketing automation soft-ware, perhaps it is your customer launching her first successful campaign.

Recurring Value

But getting customers to first value is not enough to retain these cus-tomers long term. A single action does not a habit make. In retail mar-keting, we find that 70 percent of first-time buyers never come back.Perhaps they did not have a good experience or, more likely, they simplyforgot about you. The same is true for many business products, includ-ing software. There usually is a honeymoon period just after purchasewhere users are very excited about trying out a new product or tool.However, if using the new software does not become a habit within thefirst months, much of business software ends up as “shelfware.” Shelfwareis software that was paid for but never used.

It is not always apparent that companies have lost these customers.If you are a magazine, you may only discover your subscribers havestopped reading your magazine when it comes time to renew. However,by then it is too late to create a habit of reading your publication.The same is true for other products.

In retail, customers are usually not considered lost or lapsed untilone year after their first purchases. However, our research shows that formost products, if a repeat purchase does not happen within the first fewmonths it probably will never happen. This window of recurring valueis different for each company. If you are selling cars, the repurchase timeframe could be years. If you are selling pet food, it is probably a monthor less. In flash sales on fashion sites, it is weeks.

Whatever the length of the honeymoon period, just after a cus-tomer’s first value experience with your brand, you have a lot of goodwilland an unique opportunity to convert this customer from a one-timebuyer to a repeat buyer. Your job is now to deliver recurring value.

If you get a retail customer to buy a second time, the churn rate dropsfrom 70 percent to 30 percent. In other words, whereas only 30 percentof one-time buyers come back to buy again, a whopping 70 percent ofsecond-time buyers will return for another purchase.


Predictive algorithms help predict the repurchase window for yourcompany. They can help you drill down to predict the repurchasewindow by clusters or by individual customer.

Let’s say that you determined the honeymoon period for your com-pany is two months and if a second purchase does not occur after twomonths, it will likely not happen. There are a couple of actions you cantake with this information.

First, provide helpful information to the customer immediately afterthe first purchase and offer a thank you note for the initial purchase.This is also a good time to make recommendations for what the customermight want to purchase next. We speak more about recommendationsin Chapter 10.

Second, if your product is something that is typically replaced peri-odically, such as face cream, water filters, or pet medications, send afriendly reminder a week or more before their supply is likely to runout. Most customers will find value in the reminder itself: They willconsider it good customer service that you send this reminder. Wakingup one morning to discover that you ran out of face cream or pet foodis no fun, so a courtesy reminder is very welcome to most consumers.

If after 60 days no repeat purchase has occurred, it is time to step upyour game. You have nothing to lose at this point, so you might want toconsider creating additional incentives to get the customer to buy again.

In business marketing, you similarly want to see recurring use ofyour product within the first months after purchase. For instance, if youare selling marketing automation software, you want to see customersset up a second and third campaign or landing page shortly after theirfirst. Declining usage or nonusage is a sure sign your customer may notrenew at the end of their subscription period.

New Value

Even recurring value and repeat purchases are not the end of the cus-tomer journey. If you have established a good pattern of recurring value,you may not lose this customer any time soon, but you are still leavingvalue on the table.

In Chapter 5 we talked about the difference between predicted life-time value and upside lifetime value. Predicted lifetime value accountsfor the future revenues you can expect from customers if they continue


to buy from you, as they do now. We gave the example of a customerbuying hockey tape from a sports store and the retailer trying to marketcomplimentary products to her that she is currently buying elsewhere inorder to capture new value.

If a customer is already going to the Disney theme parks every year,new value for Disney would be to get her to go on a Disney cruise vaca-tion next. In business marketing, new value means getting customersto pay for more features of the same product, or buying more productsfrom the same vendor. In our marketing automation example, new valuecould be going beyond installing simple email campaigns and configur-ing on-site personalization. If you are not experiencing the full value thata product or vendor has to offer, your chances of churn are higher. Onmany of our customer studies, we’ve found that given everything elsebeing equal, customers with the higher number of categories or dis-tinct products used, have higher future value. Engagement over multipleproducts is a strong indicator of future value.

Life Cycle Marketing Strategies

Now let’s look at some high-level strategies to give customer value inorder to get customer value. Figure 7.2 gives you an overview of lifecycle marketing strategies. We will discuss these strategies at a high levelin this chapter. For details on specific campaigns, read Chapters 11, 12,and 13.

Prospective Customer Strategies: Can We Help?

We have described many examples of what first value could mean forprospective customers. Whatever it is for your business, experiment withdifferent ideas and measure the results. Measuring results also means lis-tening to your non-buyers. Perform an exit poll on your website, orinstitute loss interviews for all lost deals in business marketing. Monitoryour prospects in whichever way you can. Using the behavioral datasuch as email signups, form fills, web browsing behavior, keywords used,marketers need to predict and trigger campaigns to prospect to con-vert efficiently to become customers. Campaigns based on behavioraldata make the experience personal and increase conversions significantly.For example, a prospect who searched for “red shoes” on the website after


Life CycleSegment Definition

MarketingObjective

MarketingStrategy

Prospectivecustomer

Somebody who hasnever boughtfrom you before.

Get the prospectto convert toa payingcustomer.

Deliver first value or firstwow.

New customer A brand newcustomer withinthe first 90 daysof her firstpurchase.

Get your newcustomer tocome backsoon and buyagain.

Welcome new customers, getcustomers to recurringvalue. Understand whythey became a customerand whether and how theyare different from existingcustomers.

Repeat/activecustomer

Somebody who hasbought at leasttwice from you.

Keep thesecustomersengaged, getthem to referothers.

Continue to delight,introduce new value. Findways to encourage them torefer friends. Introducethese customers to newcategories or services thatthey haven’t engaged withyet.

At-risk/Inactivecustomer

Hasn’t bought(or used) productor service for 90days or not highlylikely to buy.

Reengage thesecustomers.

Investigate satisfaction; give areason to try again.

Lapsedcustomers

Haven’t bought in ayear, or very lowlikelihood to buy

Reactivate thesecustomers.

Give a reason to try again.

Figure 7.2 Overview of Life Cycle Marketing Strategies

signing up for the email newsletter should get best sellers in red shoes.These triggered campaigns are highly relevant and therefore effective.

Zendesk, a provider of cloud-based helpdesk software, generatedmost of its new business through a free self-service trial. BecauseZendesk doesn’t speak to its prospects in person during the buyingprocess, it decided to monitor exactly what features free-trial userstouched. Using this information, it was able to identify which parts ofthe software customers were not using and reengineer the onboardingprocess. These changes resulted in a more than 100 percent increase inengagement during the trial period.


It is also important to be helpful during the buying process.Many customers come into Apple retail stores with no intent to buy.Once children begin playing with free games and adults start experi-menting with new phones and computers, it’s only a matter of timebefore store associates ask customers how they can help. Before long,a customer might ask about the cost of upgrading a phone and a salemight ensue.

The same can be done with digital marketing. If you see that peopleare visiting your website but not completing a purchase, stretch out ahelpful hand. The digital equivalent of “can we help” could be an onlinechat box, a pop-up, or a friendly reminder after visitors have alreadyleft your site. We speak more about remarketing campaigns such as cartabandonment, email, or display retargeting in Chapter 11.

In business marketing, you may want to pick up the phone and calla prospect with a simple “can I help.”

New Customer Strategies: Thank You

We have established that if you don’t convert the one-time buyer quickly,you won’t ever convert that customer. So the actions you take in thefirst days, weeks, and months after you have acquired a new customerare extremely important. A good place to start with new customers is asimple thank you.

Nonprofits are masters at sending thank you letters. By includinginformation about the donation’s impact, nonprofits make the donorfeel good and ensure future donations. The SmileTrain is a nonprofitthat funds cleft lip operations in developing countries. When youdonate to the SmileTrain, the organization responds by sending beforeand after pictures of a child who was physically transformed throughyour donation.

So few companies do a good job at a proper thank you that it iseasy to stand out and make a big impression. When you order fromMoosejaw, a retailer that sells outdoor gear and apparel, your order arrivesin a box with a sticker that says: “Sealed with a kiss by: Matt” and thehand-written name of the packer who closed your box. You can findmany pictures online of customers who were so impressed by this gesturethat they posted praise about it online. The boxes also come with otherstickers such as “No knife. Use teeth.” and “Don’t be surprised if you


have seen this box before. We recycle.” Moosejaw’s CMO Dan Pingreesays this allows the company to deliver customer experiences that peopleremember, not just products or transactions.

After your thank you, make sure to provide new customers withhelpful hints on how to use or maintain their new products. This couldbe a list of tips on how to use, wash, or maintain the product. Perhapsa customer service representative could even pick up the phone and askthe customer if everything is okay with a first purchase.

We have found the same is true in business marketing. If a customerexperiences problems with your product, nothing is more powerful thana customer service representative reaching out proactively. If you knowyour product is difficult to use, perhaps offer customers a free course orregular phone calls with a customer service representative.

Beyond getting customers to use your product or to come back andbuy more, postpurchase is also a good time to make upsell or next-sellrecommendations. Home improvement stores found that shortly aftercustomers bought wood to build a deck, they were in the market fora new barbeque or grill. And shortly after customers bought a newwood-pellet grill, they typically would buy grill accessories, a grill cook-book, and refill wood pellets. Predictive analytics can help you find thesecorrelated products and help you make the customer journey in greaterdetail. Once you know what products are usually bought next, you cansend customers recommendations proactively.

If you fail to convert a customer during the honeymoon period,don’t give up. Instead, step it up. You have nothing to lose so you mightas well provide a more aggressive offer. The longer you wait, the loweryour chances of bringing the customer back. You still have a much betterchance than to wait for the customer to lapse.

Repeat/Active Customer Strategies: We Love You

There is a popular misconception that if customers are actively usingyour product or service and coming back to buy time and time again thatyou should leave them alone. Nothing could be further from the truth.There is still much to lose and to gain from these customers. You alwayswant to stay top of mind with customers. There are so many brands outthere that it is easy to be forgotten. Once out of sight, it is likely thatyour brand will also be out of mind. Plus, repeat customers often still


have upside lifetime value if you can get them to buy other productsfrom you. If nothing else, repeat customers are also prime candidates tobecome brand ambassadors for your brand and to refer their friends.

A simple campaign to appreciate your best customers can be power-ful. For instance, when LinkedIn reached 200 million users, the companysent an email to its top 1 percent, 5 percent, and 10 percent most-viewedprofiles in different geographies. Many of the top users proudly postedthe digital letter they received online, and the simple gesture became oneof the biggest viral marketing campaigns in LinkedIn’s history.

Likewise, in the early years of Amazon, the e-commerce retailersent coffee mugs to loyal customers around the Christmas holidays.We received such a mug and 15 years later remember it well enough towrite about it in this book!

For repeat customers within customer life cycle management, cre-ating treatment strategies would involve understanding customers froma value, behavior, engagement, product, and brand perspectives, similarto what we outlined in the chapter where we discussed clustering.

Inactive Customer Strategies: Remember Me

Don’t give up on inactive customers. Just because they haven’t boughtanything for a while doesn’t mean they won’t ever buy anything fromyou. The first thing to do is to investigate if there is a specific rea-son why this customer stopped buying or using your product or ser-vice. You can do this by picking up the phone or via an email survey.Don’t make assumptions. One vendor we do business with was trying tobe ultra customer-centric by proactively unsubscribing consumers fromtheir email list if they hadn’t opened these emails for a while. Many cus-tomers who were unsubscribed automatically were upset. Even thoughthey hadn’t bought for a while, many said they still enjoyed reading thenewsletter and had every intention of buying again.

The sooner you act on inactive customers, the better. Don’t wait forthem to lapse. Offer constant gentle reminders of the great past expe-riences they have had with your brand and provide a gentle nudge tocome back again. Just like a simple “thank you” and “can I help” goes along way, a simple reminder can work wonders for inactive customers.

Music app Shazam found the best way to re-engage subscribersthat hadn’t logged into the service for a while was to send personalized


recommendations for songs or concerts via email or mobile pushnotifications.

If recommendations don’t work, send reminders with a simple“we miss you” and start to give your customers specific reasons tocome back. For consumers, perhaps a discount will get them to buyagain. For businesses, perhaps you offer up a free tune-up or training.The messaging and offers for inactive customers and lapsed customersare very similar, but you should increase the value of the offers as timegoes on. It will become more and more difficult to convince customersto come back.

You should take into account the personalized repurchase windowwhen doing this. Certain products have a longer time between two pur-chases, and certain customers have a longer time between purchases.No point to remind me every month to come back if you know frommy past behavior that I am the type of shopper who comes just twice ayear to stock up.

Lapsed Customer Strategies: We Miss You

Lapsed customers are those that haven’t bought for more than a year orthose that let their subscription or service lapse. Here the guiding light is“what have we got to lose?” These customers may or may not respondto your emails, but there is a greater chance than with inactive customersthat they are tuning out your communications. A strong offer is moreimportant here than personalized recommendations and reminders.

Whereas most of the strategies of reminders, escalating offers of helpand reasons to come back, and simple “I miss you” still apply for thiscustomer segment, you may need to try different channels to reach thesecustomers. If a customer hasn’t opened an email from you for a year,it doesn’t help to continue to send emails. Instead, you may want totry direct mail or Facebook custom audiences advertising to reach thesecustomers. Perhaps the customer will actually look at the postcard offerand be inspired to return to your store.

PetCareRx sent postcards to reactivate lapsed customers. Postcardsare expensive so they only sent these to customers with a relatively highlikelihood to buy. Also, the incentive for coming back to PetCareRx wasdifferent for customers with different lifetime values. Those with higherlifetime values received a greater discount on their next order.

CHAPTER 8

Play Four: PredictCustomer Value andValue-Based Marketing

The days of one-size-fits-all customer service are long gone.Not all customers are going to be as valuable to you as others.

For instance, the costs incurred from customers who frequently returnthings they purchase could outweigh the revenue from those customers.In Chapter 4 we defined customer lifetime value in detail. In this chapterwe look at strategies to segment and target customers by their lifetimevalues, a practice called value-based marketing.

Value-Based Marketing

Any business will have those low-value customers, as well as mediumand high-value customers. The trick is identifying which customers fitinto which value buckets and crafting differentiated marketing and ser-vice strategies based on the value of each customer. That means keepingperks like unlimited free shipping and returns for high-value customers,rather than the low-value heavy returners.

115


Figure 8.1 summarizes the three key strategies depending on cus-tomer value:

• High-value customers: Spend money to appreciate and retain thesecustomers. Pay close attention to retention metrics here.

• Medium-value customers: Upsell to migrate these customers to max-imize their potential. Pay close attention to upside potential of thesecustomers.

• Low-value customers: Reduce your costs of servicing unprofitablecustomers.

For simplicity purposes we broke down the customers into threesegments, high, medium, and low value. We usually designate the top10 percent of customers as ‘VIP customers’, since there should be veryfew VIP customers to pay attention to, the next 60 percent of customersas ‘average’ and the bottom 30 percent as ‘low profitability’ customers.This is easily done by ordering customers from highest to lowest revenueor profitability and choosing the top 10 percent, the next 60 percent,and the lowest 30 percent. The reason why we don’t do this based on anabsolute revenue or profitability break (for example, all customers whospent over $500 per year is in high value bucket). The reason is to alwayshave the same proportion of customers in the value segments, and track

Figure 8.1 Value-Based Marketing Strategies

Play Four: Predict Customer Value and Value-Based Marketing 117

the average value of these segments. This way retention metrics can becalculated for the same portion of the population, and this approach isproofed against change in segment averages.

The reason for choosing this 10/60/30 is more of a business choicethan mathematical. You can also choose to find the right split bylooking at the histogram of value and decide on the right split in amore mathematical way. At an apparel retailer we worked with, withinactive customers, high-value customers spent $600 on average, whereasmedium-value customers spent $120, and low-value customers spent$30. This is not atypical; in many cases we’ve seen top 10 percent high-value customers contribute close to 30-40 percent of all profits, medium-value customers contribute 60-70 percent and low-value customerscontribute anywhere from 0-10 percent.

For any mix of customer value segments, you need to pay atten-tion to how the mix changes over time, making sure that retentionand acquisition for each of these segments trends favorably. The con-cept behind value-based marketing is to understand the mix of customervalue over time. Figure 8.2 shows how this works by looking at the valuebreakdown of customers over the last 12 months (or if you’re using pre-dictive metrics, you would use predicted next 12-month value againstlast 12-month value) and cross-tabulating this against the prior 12-monthvalue status of the customer. For example, if a customer was a high valuein the prior 12 months and hasn’t placed an order in the last 12 months,this is defined as a lapsed customer. However, there are three lapsed cus-tomer segments as shown, ranging from a high, medium, to low valuesegment. High-value customers lapsing is much worse than low-value

Figure 8.2 Value Transition and Definition of Value Segments


customers (which you might even be better of without in some caseswhere low-value customers contribute negatively to profitability).

The transition matrix shown in Figure 8.2 can be used to calcu-late many useful metrics. The seven segments depicted in the figuredescribe important patterns in your customer data. Segment 1 describescustomers who have been inactive for a long time (24 months or morein this example). These are not only lapsed customers, but customersyou failed to reactivate. Most marketers have a growing number of thesecustomers over the years and provide a pool of opportunity to reactivatefrom the past. Segment 2 is customers who have existed in your cus-tomer database and were inactive and recently have been reactivated. Theimportance of reactivation is that it counterbalances efforts of lapsed cus-tomers. Segment 3 is customers who stay within their segment of valueover time. Segment 4 is customers who lapsed in the recent period whoused to be active. Segment 5 is customers who are migrating upward invalue, which shows they are increasing their loyalty and value. Segment 6is the opposite of 5; these are customers who are migrating downward invalue and signal attrition risk. Segment 7 is customers who are recentlyacquired and which value they have or are projected to have.

As we mentioned before, you can either use actual historical valueor predicted future value when utilizing this framework.

Figure 8.3 shows an example of this framework. For example, wecan see that 1,000 customers who used to be high-value customers havelapsed. It also shows that of the 21,000 customers acquired, 3,000 ofthem were high-value customers.

Last 12 months

Prior 12 months

No.Orders

HighValue

MediumValue

LowValue

NewCustomer Total

No. Orders 20,000 1,000 14,000 20,000 55,000High Value 1,000 15,000 2,000 1,000 3,000 22,000Medium Value 9,000 5,000 96,000 3,000 14,000 127,000Low Value 2,000 2,000 18,000 38,000 4,000 64,000Total 32,000 23,000 130,000 62,000 21,000 268,000

Figure 8.3 Example of Value Transition Framework


Reactivation New – LapsedNet

ChangeSpend/

CustomerValue Gain/

Loss

High Value 1,000 3,000 (1,000) 3,000 $600 $1,800,000Medium Value 9,000 14,000 (14,000) 9,000 $120 $1,080,000Low Value 2,000 4,000 (20,000) (14,000) $15 $(210,000)Total 12,000 21,000 (35,000) (2,000) $2,670,000

Figure 8.4 Net Loss/Gain from Acquisition, Reactivation, andLapsed Segments

In Figure 8.4 we show a practical usage of this framework for track-ing performance of a customer base. We know that for every lapsedcustomer, we can either acquire a new one or reactivate an existingcustomer. This framework also deals with an important factor; are weacquiring and reactivating customers of same value compared to lapsedcustomers. This shows the retention in a more granular way to uncoverwhat we call silent attrition, that is, we could be acquiring/reactivatingthe same number of customers, but of lower value, therefore losing enter-prise value. In the example given in Figure 8.4, it shows that the activecustomer count (total net gain/loss, which is -2,000 customers) wentdown; however, due to increased retention of higher value segments,the customer value increased by $2.7 million.

Retaining High-Value Customers

Until recently, many firms were unable to identify their high-value cus-tomers, let alone give them the white glove treatment. Although airlines,banks, and casinos know it pays to make big investments in retentionincentives for high-value customers, too many midsize organizations stillignore their best customers.

Spending to retain high-value customers pays off. Often a small per-centage of customers make up the majority of revenues. A cosmeticsretailer we work with found that 50 percent of revenue came from just20 percent of customers. When analyzing its best customers, a popu-lar flash sales site found that some of its best shoppers spent more than$100,000 a year with the retailer.

When a popular home improvement website first started to calculatelifetime value for their customers, it was surprised to find that some


customers spent more than 20 times more than the average customer.These so-called whales were so important to the company that the CEObegan picking up the phone to get to know these customers one by one.From those conversations, new ideas emerged on how to better serve andattract more high-value customers. Similarly, a flash sales site decided tosend a box of chocolates for Christmas to their entire top 1 percentageof customers. This was well worth the money as their top 1 percentagemade up 20 percentage of their revenues. Technology companies Wufooand Stripe are well known for sending handwritten notes to wow theircustomers.

In the Harvard Business Review article “Manage Marketing by theCustomer Equity Test,” authors Robert C. Blattberg and John Deightonrecount the experiences of McDonald’s. The corporation’s managersnoted that the value of what they call super-heavy users—typically malesaged 18 to 34 who eat at McDonald’s an average of three to five times aweek—account for a whopping 77 percent of its sales. Naturally, retain-ing these customers and getting them to eat at its restaurants more oftenis a priority. As a general rule it is much easier to get a current customerto use you more often than it is to get a new customer.

Another example of value-based marketing is airline loyalty rewardsprograms that are based on the dollars you spend with the airline, ratherthan the miles you have flown. This way, higher paying customers auto-matically get a greater reward than lower paying customers.

Marketers should use their retention budgets in order to proactivelyretain high-value customers. If you have an accurate projection of futurecustomer lifetime value, you can experiment on what it takes to retainthis customer. Some organizations might consider crafting separate mar-keting plans or even build separate marketing teams to focus on acquisi-tion and retention efforts. Titles containing “customer marketing” and“customer success” are increasingly popular across many verticals.

We discuss retention strategies in greater detail in Chapter 13.

Growing Medium-Value Customers

Your main strategy for medium-value customers is to nudge them intothe high-value customer bucket. In addition to selling these customersmore of the products they are already buying, start predicting what otherproducts and categories these shoppers might like. A customer than


who spends $1,000 across three categories, say furniture, clothing, andkitchenware, has more future value than a customer who spends $1,000on furniture only. Your strategy should always be to try and get customersto buy from you across multiple categories.

Customers who already like a company’s products and services areless expensive to serve with new products and services. For that reason,you should be recommending different products to existing customers,as well as adding new products and services to meet your customers’evolving needs.

Companies that fail to use their knowledge of customers to promoteor develop the products or services those people will need next areleaving the door open for another company to lure those peopleaway. Although it is tempting to use new products to win whole newmarkets, it almost always makes better sense to stick with existingcustomer segments.

Mavi Jeans has been very successful with migrating customers tohigher value segments. For example, the company found a segment ofjean lovers who also had an affinity with certain tops but had never beenmarketed these tops. Adding campaigns based on these insights increasedcustomer lifetime value by 36 percent.

In his Harvard Business Review article on loyalty-based management,Frederick F. Reichheld tells the stories of Entenmann’s of New York,a leader in specialty bakery products, and the car company Honda.When Entenmann’s saw its sales leveling off, it discovered that as its corecustomers aged, they were looking for more fat-free and cholesterol-freeproducts. So instead of trying to find new customers for its existing prod-ucts, Entenmann decided that it was more economical to launch a newfat- and cholesterol-free line of products to serve its existing customers.The new line has been extremely successful. Similarly, Honda figuredout how to increase a Honda’s owner repurchase rate to 65 percent,versus the industry average of 40 percent, by launching new cars thataccommodate its customers’ evolving needs. For example, it successfullymarketed the Accord wagon to previous owners of the subcompactHonda Civic. The wagon was designed to accommodate the evolvingneeds of customers acquired in their early twenties who were nowmarried with children and needed a larger car.

We discuss strategies to grow customer value in detail in Chapter 12.


Reducing the Costs to Service Low-Value Customers

Every brand will find itself with a segment of unprofitable customers.It is best to try to reduce the costs to service low-value customers,rather than firing these customers outright and leaving the door openfor competitors to gain strength. In any business scenario, this is a fact oflife. When acquiring customers, it is impossible to reject low-value cus-tomers. However, keeping it in check, that is, understanding, measuring,and monitoring the size and value of this segment over time, is extremelyimportant. Once you know who is in this segment and why, you canfocus on formulating strategies to reduce the cost to serve/market tothis segment. Fortunately, it is now possible to differentiate the level ofservice offered to different customer segments.

A fashion flash sales site found a particular customer segmentreturned more clothes than they kept and therefore were very unprof-itable. The company then decided to market only jewelry to thiscustomer segment because online return rates on necklaces and earringsare much lower than on clothes and shoes. Other retailers have startedto differentiate by offering free shipping to some but not all customers.If you return a lot of products, you will likely not receive the freeshipping privilege in the future.

To that end, First Union Bank has a system called Einstein, whichautomatically assigns a green, yellow, or red flag to each customer.Service representatives are instructed not to waive fees for red customers,waive them for green customers, and use discretion for yellow cus-tomers. They have generated an estimated $100 million in incrementaladditional annual revenue based on this differentiated strategy.

Likewise, First Chicago Corporation imposed a $3 teller fee onmoney-losing customers (3 percent of their customer base). By identi-fying loss-making, high-volume users and sending them notices of feeincreases, the company reduced its 11 million customer base by 450,000in six months.

CHAPTER 9

Play Five: PredictLikelihood to Buyor Engage to RankCustomers

Propensity models, also called likelihood to buy or response models, arewhat most people think about with predictive analytics. These mod-

els help predict the likelihood of a certain type of customer behavior,like whether a customer that is browsing your website is likely to buysomething. In this chapter we examine how marketers can optimize any-thing from email send frequency, to sales staff time, to money, includingdiscounts, when armed with information about likelihood to buy orlikelihood to engage.

Online pet pharmacy PetCareRx has served pet owners for morethan 15 years. It sells many products that customers have to reorderat varying times from 3 months to 12 months. Like most retailers,PetCareRx took a one-size-fits-all marketing approach, offering a setcalendar of discounts and promotions to all customers. But not allcustomers are alike and many are looking to buy at different times ofthe year. Using predictive analytics, PetCareRx was able to differentiatediscounts across customers, leading to higher sales and retention withoutincreasing costs. Customers were ranked according to their likelihood

123


to buy. Based on that ranking, PetCareRx was able to determine whichdiscounts would obtain the optimal response from each customer andoffer minimal discounts through email or mailed postcards to customerswho were already deemed likely to buy and offer larger discounts forcustomers who were less likely to buy. The surgical promotions droveincremental margin from customers who were already motivated to buyand incremental revenue from customers who previously felt no incen-tive to buy. Thanks to this and other predictive marketing campaigns,quarterly sales increased by 38 percent from the year before, profitrose by 24 percent, and customer retention increased by 14 percent.Plus, the changes allowed PetCareRx to more than double its cam-paign response rates without increasing the marketing or promotionalbudget by a single dollar.

Likelihood to Buy Predictions

To predict which prospects are ready to make their first purchase, a like-lihood to buy model evaluates nontransaction customer data, such ashow many times a customer clicked on an email or how the customerinteracts with your website. These models can also take into accountcertain demographic data. For example, in consumer marketing theymay compare gender, age, and zip code to other likely buyers. In busi-ness marketing, relevant demographics may include industry, job title,and geography.

Here’s how it works: the models compare the prepurchase behaviorof prospective buyers to the prepurchase behavior of thousands or mil-lions of previous customers who ended up buying, comparing attributeslike what emails they opened and what products they spent the mosttime looking at. The prospects that behave most like the previous buyersare tagged as “high-likelihood buyers” and marketers can then alter theway they interact with them to increase the likelihood of closing a sale.Once you’re armed with this data, you can prioritize your investment ineach prospective customer.

Likelihood to Buy for First-Time Buyers

For consumer marketers, likelihood to buy predictions allow you todecide how much of a discount you might allocate to a certain customer

Play Five: Predict Likelihood to Buy or Engage to Rank Customers 125

because people who are already more likely to buy won’t need as aggres-sive of a discount as customers who are less likely to buy. The models thenget better over time, as companies collect more data and automaticallytest whether predictions actually become reality.

For instance, the large European household appliance manufacturerArcelik maintains a call center where employees are given a list of cus-tomers who are likely to be ready to buy a new washing machine withinthe next few months. Agents then make calls to these customers withoffers such as a year of free detergent with the purchase of a washingmachine. The tactic works well for considered purchases, such as refrig-erators or cars, and larger-ticket items such as high-end fashion apparel.

A high-end shoe brand provides store associates with lists of cus-tomers to call too. The store associates have already developed strongrelationships with their customers but they can be even more successfulwhen armed with predictive analytics. Employees can now see whichcustomers are likely to be interested in a certain style when a new sea-son’s shoe comes out, based on customers’ past behavior or how similartheir purchase habits are to other customers. Employees can then reachout to customers with that information. A call could go something likethis: “Hi Joe, it has been a while since we’ve spoken. I just wanted to letyou know that there is a new cross-country running shoe I think youmight like. It’s similar to the shoes you bought two years ago, but in anew material. I have put a pair aside for you in your size. If you havetime, perhaps you could stop by on your way home from work to have alook?” Who would not want to receive a call or an email like that fromtheir personal shopper?

As reported by the New York Times and others, President BarackObama used propensity models, specifically propensity to vote for theDemocratic Party, to help him win reelection in 2012. His staff of vol-unteers could not possibly meet with every voter in the country so thechallenge was to find the undecided voters. There was no point spend-ing time or money trying to woo diehard Republicans who would notchange their minds anyway, or diehard Democrats who were alreadylikely to vote for Obama. Rather, using propensity models, Obama’steam of data scientists found those voters who were undecided but couldstill be persuaded. They then focused on finding already strong Obamasupporters in the undecided voter’s social circle and asked them to spendtime with the undecided voter to explain their views.


Likelihood to Buy for Repeat Buyers

What good is spending money to acquire new customers if they onlybuy once and do not return? Therefore, it is not only important to pre-dict likelihood to buy for first-time buyers, but it is equally importantto predict likelihood to buy for repeat buyers. Your goal is to keep cus-tomers coming back time and time again. It is happy and loyal customerswho have a large lifetime value, and many customers with a large lifetimevalue make for large revenues and profits for your company.

Predicting likelihood to buy for repeat buyers is a lot easier thanpredicting likelihood to buy for first-time buyers because there is a lotmore information to go on. The likelihood to buy model for repeat pur-chases evaluates earlier transactions as well as other interactions similar tothe model for prospects. However, the added information derived fromthe first purchase can significantly improve the accuracy of the likeli-hood to buy model for repeat purchases, as compared to a similar modelfor prospects. Unlike the first purchase predictions, repeat purchase pre-dictions utilize all interactions of the customer, such as past purchases,returned purchases, and phone calls to customer service.

Choosing the Right Level of Discount UsingLikelihood to Buy

There are two main uses for likelihood to buy predictions: whichcustomers to focus on and how much money, including discounts,to spend on each customer.

Choosing your audience carefully is important if you want to opti-mize marketing ROI, because reaching customers can be expensive. Forinstance, a direct mail or call campaign that costs $1 per customer inter-action and has a 2 percent purchase rate could cost $50 a person beforea discount is even given. If you can target your audience and make thecommunications more relevant, the purchase rate could be significantlyhigher, say 10 percent, reducing the cost to reach each buyer significantly.In addition to choosing the right people, you can increase purchase ratesby including relevant recommendations or content and to communicatewith customers at the right time. Focusing on relevancy will allow com-panies to reduce their dependency on high discounts. Using this strategyit is possible to significantly reduce discounts as a part of their customeracquisition strategy. Figure 9.1 shows how a large U.S. retailer was able


Figure 9.1 Percentage Customers Acquired with Discounts

to reduce the number of customers lured by high discounts from 36percent to 27 percent, well below the industry average of 31 percent.

Discounts and other incentives can still be used as a sweetener in nec-essary cases, such as targeting a customer who has abandoned her cart.However, we recommend you do not give out discounts to everybodyor else you will train your customers to get used to discounts. Instead,start with recommendations and reminders and use discounts only if youneed them. The lack of understanding of the individual customer leadsto blanket discounts, which reduces overall profit margins considerably.Only about 20 percent of customers are “discount junkies,” people whowill only make purchases when given a discount, according to an anal-ysis of 150 retailers. About 15 percent generally pay full price for mostof their products, while the majority of customers fall somewhere inbetween, the data show. By analyzing their customers’ behavior, mar-keters can determine which customers may need more encouragementin the form of a gift or discount. It can also help marketers flag cus-tomers who will come back and buy anyway, so no additional monetaryencouragement is needed. This model helps to maximize both revenueand profitability from each member of your customer base.

Targeting discounts is good for business and good for customers.By surgically targeting discounts, retailers avoid margin erosion, which


in turn reduces the price increases retailers would normally have to taketo make up for the lower profit margins. That practice effectively lowersprices for all customers. This is a simple but very powerful paradigm shiftfor marketers who traditionally focus on product lines, merchandising,and one-size-fits-all discounting. Behavioral and life cycle data availableat the customer level combined with the predictive insight about likeli-hood to buy allows marketers to surgically discount to maximize marginand customer lifetime value.

Also, if you are going to include an incentive or discount to acquire,retain, or reactivate a customer, there is no need to go beyond the levelof discount that customers are used to seeing. Predictive analytics allowsyou to tailor offers based on what level of discount triggered past cus-tomer purchases. Customers with a high likelihood to buy may get alower discount but perhaps other perks such as early access to productsto drive more purchases. Customers with a low likelihood to buy orthose who only buy on discount (a behavior-based cluster) might get ahigher discount.

A multichannel retailer of sports goods used to send all of itscustomers a 50 percent clearance discount at set times during theyear. This campaign was motivated by its desire to clear old inventory.In other words, it was a product or merchandising-centric campaign,not a customer-centric campaign. This campaign trained customers towait for the big annual sale and brought in few dollars for the retailer.Using predictive analytics, this company analyzed the likelihood to buyand discount sensitivity of all of its customers. If a customer is likelyto buy with an offer of 25 percent off, you don’t need to give them a50 percent discount. Using predictive analytics, the retailer identified theright discount level for different groups of customers. Now the retailerstill sends discount emails at set times, but it will send different levels ofdiscounts to different groups of customers. It sends just enough to getthe customer to buy, but not too much to give away unnecessary margin.Using this approach the retailer was able to achieve an overall increase inrevenues of 20 percent.

Predictive Lead Scoring for Business Marketers

If you are in business marketing you can equally prioritize your invest-ment using likelihood to buy models. You can ensure that your sales


team spends most of their time with the prospects that have the highestlikelihood to become buyers. This can have an enormous impact.

Consider the following example in marketing business software.Let’s say you have a free trial for your software. Not all people who signup for your free service have a serious intent to buy. It is not uncommonthat 70 percent of free-trial signups are made just out of curiositywithout an immediate need or budget to buy. Another 20 percent areserious evaluators and 10 percent are on the fence: they could go eitherway. If you randomly call people from a list, perhaps based on companysize, it is easy to see how you may end up wasting your entire day onprospects that aren’t serious. It is especially important to prioritize yourtime if you are serving small- and medium-size businesses, of whichthere may be millions.

The main difference between predicting likelihood to buy in con-sumer and business marketing is due to the nature of the purchase deci-sion process. In most business marketing, the decision process is longand complex. Figure 9.2 compares the considered and quick decisionprocesses.

With a considered purchase, which most business marketing fallsunder, the decision process is longer and includes many interactionsbetween the marketer and the buyer. This requires special attention tofocus on all prospects within the decision funnel. Hence business mar-keters turn to all interactions and signals from prospects to determinewho is most likely to buy and therefore worthy of time and attention.

Because the replacement and delivery cycles for vendors, deals,services, and products can take a long time, most B2B marketers arehyperfocused on acquiring new customers, rather than getting existingcustomers to come back, where likelihood to buy first purchase modelstake greater importance.

Considered Quick Decision

Decision Cycle 6 months 1–7 daysInteractions during sales cycle 10 2Average order value $30,000 $200Replacement cycle 1 year 1 monthUsers impacted Many 1Functionality Complex Simple

Figure 9.2 Considered versus Quick Decisions


Basic Point system: Each activity has specific points and these are added tocreate a score, which becomes the lead score. The score for eachactivity is arbitrarily determined. For example, a documentdownload is 5 points and an email click is 1 point.

RFM-like This is similar to Recency-Frequency-Monetary (RFM) scoring used incatalog marketing, where the leads get a composite score based onnot only activity, but also size of the opportunity, deal size, andengagement metrics such as email opens, web visits, webinarattendance, and so on.

Predictive leadscoring

Predictive lead scoring builds on the first two basic models by learningfrom the past in a statistical manner. The predictive model learnsfrom successful deals closed in the past and looks at the prospectbehavior prior to deal closing and learns from those behaviors to helpcome to a decision.

Next actionscoring

The best models not only predict a lead score but also the next actionthat could increase the overall likelihood of closing the deal.These are highly complex models that are mostly custom-built.However, these models take into account specific sequences of eventsrather than mere event occurrences.

Figure 9.3 Lead Scoring Methods for Business Marketers

Predictive models are not the only way to prioritize prospects forbusiness marketers. However, predictive models are by far the most accu-rate and relatively easy to use. Figure 9.3 gives an overview of lead scoringmethods beyond predictive methods.

Likelihood to Engage Models

Likelihood to engage models predict how likely it is that a customerwill open or click on your emails. High email engagement is a strongpredictor of buying intend. On the other hand, if likelihood to engageis low, subscribers may opt out from your list. Many consumers nolonger bother to unsubscribe from your list, but simply stop openingyour emails. For all intents and purposes this has the same effect, andcost, as a true opt out. When a customer unsubscribes from your mail-ing list, you will no longer be able to reach this consumer with pro-motions. Every retail marketer knows that email campaigns mean salesand cash for the retailer: with every promotional email sent out, neworders flood in. Email marketing works well to drive revenues. However,when it comes to email, you need to balance the short-term revenues


it can yield, with long-term revenues. If you send too many emails,you may receive short-term revenues from some customers, but lose thelong-term revenues of those who unsubscribe.

The challenge for most marketers is to send as many emails as possi-ble in order to generate as much engagement and purchases as possible,without causing a subscriber to opt out or ignore you. By using a like-lihood to engage model, you can potentially send fewer emails to eachcustomer, drastically lower unsubscribe rates, and achieve higher cus-tomer engagement. If a customer isn’t going to open an email, thereis no point in sending it. That email may seem free, but in reality optouts cost big bucks. In fact, our research shows that every opt out costsyour company about 60 percent of that customer’s future lifetime value.Let’s say a customer’s future lifetime value is $1,000, because you expectthat customer to make 10 purchases of $100 each over the course of thenext three years. Now if this customer unsubscribes from your email list,they will not be notified about new product offerings. Without an emailto remind them, they may place only four orders, instead of 10, over thenext three years. In other words, their potential future lifetime value justwent down from $1,000 to $400—a reduction of 60 percent.

When Uncommon Goods, an online retailer of unusual gifts,adjusted its email contact frequency, results were amazing. It was ableto lower its unsubscribe rates by 50 percent without hurting sales. Itachieved the same results with fewer emails, because the customerswere delighted that emails were just with the right frequency, andthe frequency adjusted automatically when customers increased theirengagement. The key to the Uncommon Goods success was to varyemail frequency by customers, depending on their level of engagementand their likelihood to unsubscribe.

In our experience, email subscribers typically fall into five sepa-rate groups based on their likelihood to engage: Enthusiasts, Mainstreet,Sleepies, and Phantoms, and one special group, the Newbies. Enthu-siasts are the subscribers that have the highest likelihood to open andclick your emails. Figure 9.4 summarizes the open and click rates foreach of these segments. The click rate of Mainstreet subscribers is abouthalf that of Enthusiasts and a fraction of the open rate of the Enthusiasts.Sleepy subscribers have very low engagement predictions, and Phantomsubscribers are ones with virtually no engagement, receiving your emailsbut rarely opening them. Newbie subscribers are a special group: they are


Figure 9.4 Open and Click Rates of Different Email Segments

the ones who have recently signed up for your email list. Their open andclick rate tends to lie in between Enthusiasts and Mainstreet subscribers.

We found that email subscribers who most frequently open and clickon emails also generate the most money—not because they buy moreexpensive items but because they buy more frequently. For every $1 spentby email Enthusiasts, Mainstreet subscribers spend $0.58 and Sleepy sub-scribers, who rarely open email, spend only $0.34. Enthusiasts bought1.75 times more frequently than Mainstreet customers and 2.9 timesmore often than Sleepies.

Interestingly, once you get a customer to buy, they buy for a similardollar value regardless of their engagement level, which can be an oppor-tunity for the savvy marketer: your email enthusiasts already love yourbrand and buy often. You can try to nudge them toward higher-valueproducts with the right incentives.

Newbies, which are email subscribers who have recently graduatedfrom their email welcome campaign, are very lucrative customers: in ourexample they spent $1.24 for every $1 spent by Enthusiasts. Newbies alsohave very high open and click rates. In this respect they can be thought ofas “baby Enthusiasts.” But they differ from Enthusiasts in one major way:they also have the highest unsubscribe rate of all segments. So Newbiesare very engaged yet very reactive: they are opening and clicking youremails but are quick to unsubscribe if they don’t like what they see.We usually see that more than 60 percent of unsubscribes from a newemail subscription come within the first 90 days of subscription. This isan important takeaway: realize that new subscribers are almost alwaysstarting out very enthusiastically. They are basically yours to lose, andalmost all behave like your best customers.


Another behavioral difference between different groups of emailsubscribers is the sales channel mix for the different email engagementsegments. Email Enthusiasts buy primarily from email click-throughreferral. At the other end of the spectrum, email Sleepies buy primarilythrough search engines or directly on your website (without referral).This suggests different behavior patterns. Customers that are less engagedwith your emails are more utilitarian shoppers: they have a need andseek you out (either through a search engine or by coming directly toyour website) and then they buy. Customers who are highly engagedwith your emails are also strongly influenced by these emails. You mightsay that they are more impulsive: they might not have planned to buy,but were tempted by your irresistible email!

How Often to Email Your Customers

Finding the correct email cadence is tricky. You are looking to strike abalance between maximizing short-term email click-through revenue(typically by sending more emails) and minimizing long-term lossdue to email opt-outs. The answer varies by customer segment: emailEnthusiasts may not mind hearing from you every day, whereas thosewho rarely open your email cannot be emailed more than a couple oftimes a month. What is at stake is for customers to unsubscribe and therisk of losing future revenues that comes from being able to email relevantoffers. As Figure 9.5 illustrates, there is a magical point that optimizes the

Figure 9.5 How to Balance Short- and Long-Term Email Revenues


net revenue from your email campaigns as deducting the cost of sendingas well as the lost, future revenues due to unsubscribes.

So how to find that magical sweet spot? No surprise here: Testing!But be careful because email frequency testing needs to be done right. Inorder to truly test the effect of reduced frequency, you must design yourtest to remove other factors that might lead you to wrong conclusions.For example: let’s say you typically send two emails per week to yourentire contact database. On Mondays you send a “hot deals” email andon Thursdays you send a personalized product recommendation email.Now you want to test a 50 percent reduction in email frequency on a testgroup. You sample your test group randomly (so far so good) and sendthem only the Monday email for one month and measure the response.The problem is that the response you get actually measures three con-founded factors: (1) reduced frequency, (2) day of the week, (3) emailcontent type.

Instead, we recommend measuring both the cumulative opt-out rateas well as the purchase rate during the test period. Figure 9.6 walks youthrough the thought process. Purchase rate is defined as the percent ofcustomers in the test group that made any purchase during the test.This is a better sales metrics than purchase dollars, which can be verynoisy. The interplay of these two metrics will guide you in your deci-sion whether to reduce frequency or not. The key is to run a reducedemail frequency test against a control group: for some of your email

Figure 9.6 When to Reduce Email Frequency


subscribers continue as you usually would. For another group reduceyour email frequency by 25 percent or even 50 percent. Run the test foras long as it takes your control group to receive at least 10 emails, but 15is even better (yes, waiting is painful).

Now compare the two groups and ask yourself: Is the unsubscriberate for the “reduced email frequency” group lower? If not, you don’tneed to change anything, because the main incentive to reduce fre-quency is to reduce unsubscribes. However, if the unsubscribe rate wentdown, the next question to ask is whether the purchase rate is alsoreduced. If the unsubscribe rate went down, but the purchase rate isunaffected, then reducing email frequency is a no-brainer. It will giveyou the same short-term revenue while preserving a lot of future rev-enue upside (by retaining these people on your email list). If, however,the purchase rate is reduced when you send fewer emails, then you needto ask yourself whether a strong case can be made that the long-termrevenue gain outweighs the short-term revenue loss. This is a very hardquestion to answer—you need to estimate the cost of an unsubscribe interms of loss of future revenue relative to the short-term email revenueyou get per customer from the extra emails you send.

Most marketers err on the side of caution and make decisionsbased on the short-term revenues. I can’t blame them. Fortunately, inour experience, some clients can reduce email frequency by as muchas 50 percent without hurting immediate revenues (purchase rate).Reduced frequency may have other benefits: reduced send cost andpotential increase in deliverability—but these are secondary consider-ations only after you convince yourself that you can reduce opt-outswithout reducing revenue.

It is possible to reduce the unsubscribe rate without hurting rev-enues. Figure 9.7 plots the trajectory of one company following thisstrategy. They were able to maintain the cumulative margin per customerover time, while at the same time reducing the cumulative unsubscriberate from 1.5 percent to 1.1 percent.

If you want to get started right now with frequency control, ratherthan waiting for the full test cycle, here are some final recommendations:

Newbies In the first 60 days from opt in do not send regular emails tothis segment—only life cycle and triggered emails—Newbies are quickto opt out if they get a lot of nonrelevant emails.


Figure 9.7 Reducing Opt-Outs with Equal Margin Dollars

Phantoms and Sleepies Send general emails at much lower fre-quency than your normal cadence. Example: if you typically send cus-tomers one email a week, this segment should only receive one emaila month. If you send three to seven emails a week, they should receiveonly one a week.

Enthusiasts Do you have the ability to produce additional content?If so, consider sending an extra email to your Enthusiasts to see if youcan get them to buy even more.

CHAPTER 10

Play Six: PredictIndividualRecommendationsfor Each Customer

ANorth American beauty and cosmetics company with hundreds ofstores was looking to personalize its communications to hundreds

of thousands of customers and ensure that every online interactionbetween the brand and its customers was consistent with its messaging.It wanted to shift the mind-set from being discount-driven to improvingcustomer service and satisfaction. The company chose to combinecluster-based targeting with personalized recommendations to sendcustomers more strategic, personalized offers. The company first usedpredictive analytics to organize its customers in product-based clusterssuch as “bath and beauty” and “face cream.” Then it emailed each ofthese customers content and recommendations that were based on theircluster. Clearly customers liked the emails and the company was able toincrease revenues per email by six times.

In this chapter we learn all about making recommendations to cus-tomers. Recommender systems have been around for almost 20 years,Amazon being the primary example that started using this early on.There are three parts to making personalized recommendations: sending

137


recommendations to customers at the right time, understanding thecontext, and sending the right content.

First generation recommender systems used simple rules configuredby human beings based on things like keywords or titles. In other words,a merchandiser or content marketer set up a rule so that everybodywho bought shoes would get a recommendation for protective sprayshortly thereafter: “if browsing for shoes, also recommend spray.”These first-generation systems used people, rather then predictivealgorithms, to make recommendations.

Especially in categories with large selections, such as books, videos,and content, recommender systems utilizing the so-called wisdom of thecrowd are more effective. Actual usage data or review data from users hasmore information than metadata like title, description, and keywords thatdescribe the content. When we are trying to find a restaurant, book, ora movie, we tend not to trust canned descriptions of the product weare looking to buy. Instead, we ask trusted friends and colleagues whatthey think. The same logic is used with recommender systems, whichcan figure out which customers are most similar to an individual userand use behavioral data (usage, reviews, purchases, views, downloads)to recommend other content for that person. This strategy will allow formore relevant recommendations, rather than just trying to recommendproducts based on certain labels or content. In mathematical terms, theseuser-based recommendations are called collaborative filtering.

Choosing the Right Customer or Segment

The first question to answer is who to make a recommendation to andwhen. Good times to make recommendations are either during a pur-chase or after a purchase—and at certain times during a customer’s lifecycle, such as when you have not heard from a customer for a while.These recommendations are referred to as upsell, cross-sell, and next-sellrecommendations, respectively.

Recommendations Made at the Time of Purchase

Upsell and cross-sell recommendations can be made to customers dur-ing a purchase, served on a website’s product page, or during checkout.

Play Six: Predict Individual Recommendations for Each Customer 139

A basic example of upselling is asking a McDonald’s customer if she wantsto supersize her meal, but similar instances can be found in all industries.You could suggest a higher-end version or a multipack of the same prod-uct, perhaps at a better price. Upsell recommendations are typically tiedto a specific product: each product has other suggested products that canbe used as upsells.

Cross-sell recommendations are also made at the time of pur-chase. Rather than recommending buying a larger or better versionof a specific product, cross-sell recommendations are made to suggestother products that are typically bought alongside this specific item.The recommendation could read: “customers who bought a printeralso tend to buy printer ink…” and you could offer a modest discountif the customer decides to buy your cross-sell bundle. Like upsellrecommendations, cross-sell recommendations also tend to be tied tospecific products: every product has suggested products that can be usedas cross-sells.

Upsell and cross-sell recommendations are a great way to increaseaverage order value. Most upsell and cross-sell recommendations are tiedto the product, rather than to a specific customer.

Of course recommendations do not have to be product recommen-dations. The online fare comparison engine Kayak developed a priceprediction tool to advise the shopper on whether they should buy orwait, depending on the confidence of a price drop. Kayak uses thiscompetitive advantage to improve customer experience and retain cus-tomers: “We want [travelers] to get to the best decision for their needsas easily as possible,” said Robert Birge, Kayak’s chief marketing offi-cer in an interview with USA Today. (Source: www.usatoday.com/story/travel/flights/2013/01/15/kayak-advice/1834225/.)

The European eyeglass retailer Alain Afflelou revolutionized theoptic market by launching the “Tchin Tchin” offer: when the customerbought a pair of corrective glasses, she could buy corrective sunglassesfor only 1 additional euro. Used as an acquisition tool, this cross-sell recommendation at the time of the purchase presented limitedupsell opportunities but enabled the company to grow its customerbase by 50 percentage in three years according to the CEO in an articleon the marketing strategies website www.strategies.fr. (Source: www.strategies.fr/actualites/marques/155836W/tchin-tchin-alain-afflelou-recidive.html.)

http://www.usatoday.com/story/travel/flights/2013/01/15/kayak-advice/1834225/

http://www.strategies.fr

http://www.strategies.fr/actualites/marques/155836W/tchin-tchin-alain-afflelourecidive.html


Recommendations Made After a Purchase

Next-sell recommendations are typically made after a customer hasalready made a purchase. This type of recommendation could beincluded in a thank you page or in the confirmation email.

The best next-sell recommendations are specific for each customerand take into account more customer data than just the most recenttransaction. By the time somebody has completed a transaction, youknow who this person is and should be able to make a more personal-ized recommendation. The more you know about a person, the betterthe recommendation. So if you can analyze all the purchases that a per-son has made, both online as well as in-store, your recommendationswill be more accurate than if you are only looking at online transactions.So, as we discussed in Chapter 3, make sure to base your recommenda-tions on complete customer profiles that tie all customer actions back tothe same person.

Remember the home improvement store that found that people whobuild decks tend to be in the market for a grill shortly thereafter? A mar-keting program was devised to capitalize on this knowledge. Similarly, thegrills company who found customers need wood pellets after their initialpurchase of a grill now sends regular pellet replenishment reminders.

Recommendations Made During the Customer Life Cycle

You can try to use recommendations to reengage or reactivate lapsed cus-tomers. In this case, first you use predictive analytics to identify a groupof customers at risk of leaving. Then you can reengage these customerswith a personalized email. The recommendation can be a product, con-tent, or relevant person. The power of recommendations is that they canbe dynamically inserted into a web page or email and create an entirelypersonalized experience without having to redo the creative for eachcustomer. The web page or email design is the same for every customer.Even the text in the page or email could be the same, telling lapsed cus-tomers “we miss you, please come back soon, we have these productswaiting for you” but include person-specific recommendations.

Be careful using product recommendations if the customer has notbought for a long time. The product recommendations might be obso-lete and the context of the customer might have changed completely:it is a new season, the customer might have picked up new hobbies, or


life events might have occurred. Make sure you base your dynamic con-tent on the latest information you have on the customer, and in somecases it might be better to use recent content, rather than products, toentice the customer.

Understanding Customer Context

Beyond the right time to send a recommendation, there is more contextto take into account when making recommendations. For example, if auser who generally likes documentary movies is trying to find a moviehe can watch with his children, the context needs to be recognized andthe recommendation made relevant for that context. Likewise, a retailermight recognize that a shopper, who usually purchases work clothes,might this time be shopping for a special occasion and make a contextualrecommendation, upselling or cross-selling jewelry and shoes to fit theoccasion. Context could also be what products a customer has boughtin the past. If you are going to recommend accessories for an electronicdevice, you better make sure that the accessories are actually compatiblewith the device the customer has purchased in the past.

Basic recommendations are “people who liked this product, alsoliked…” type recommendations. We call these product-to-product recom-mendations because a recommendation is generated with a specificproduct as the starting point or context. This could also be content-to-content or person-to-person type recommendation. By looking at whatcustomers frequently buy or read together, you can make recommen-dations even if you do not know anything about the person looking atthe page. These recommendations are often added to a product page.While you are looking at a specific book, you also see the other bookspeople who bought that book liked. If you are looking at a person’sprofile on LinkedIn, you may receive recommendations for other profilesto check out. If you are reading an article on your favorite news site, youmay be recommended other content to check out.

The problem is that perhaps you have already bought some of thebooks that are being recommended to you, or read some of the arti-cles that are being displayed. Also, there may be very different personasor clusters considering the same product. If a school teacher and a stu-dent both look at the same book, they may have very different reasonsor interests in considering the purchase. This is where person-specific


recommendations come in. If you also know a customer’s demographicprofile, past behavior, and location, you have more personal context tomake accurate recommendations about companion products. We callthese user-to-product recommendations because the starting point for makinga recommendation is the information you have about a specific per-son. Person-specific recommendations require you to both recognizethe customer as well as have a history with this customer rich enough togenerate recommendations.

Making irrelevant or out-of-context recommendations is probablyworse than not making any recommendations at all. A well-knownForrester analyst received an email promising her the latest shoes in justher size. She eagerly clicked the email link but was redirected to a pagewith oversize shoes (rather than her size X). She was so disappointedabout the experience that she tweeted about it. A happy customer tells5 friends, but an unhappy one tells 20! So make sure your customer datais complete and accurate before you start making recommendationsbased on customer profiles.

Also, have you ever received an email or an advertisement with rec-ommendations specifically for you, only to find that when you clicked onthe email or advertisement it took you to a generic web page? Here theretailer is offering the promise of personalization but then does not fol-low through. The problem here is coordination across channels. Clearlythis company’s email and web systems are not coordinated. Again, anincomplete personalized experience may deliver more disappointmentthan delight.

As we mentioned in Chapter 7, the freemium business modelrelies heavily on the use of relevant and contextual recommendationsto transform active users into paying customers. The online musicplatform Spotify manages to keep a steady ratio of 25 percentagepaid to free users (15 million paying users in January 2015), evenafter the early adopter wave. In an article on TechCrunch the com-pany’s top management team explains that their mobile applicationdrives much of this growth. Indeed, in Spotify’s mobile free version,users can listen to an artist, but they have to listen to advertise-ments regularly and they cannot pick a specific song. If they try todo so several times in a row, a recommendation of the premiumversion at $9.99 per month pops up in the application. Spotifymakes sure to emphasize all the advantages of the premium versionwhen the user’s frustration reaches its paroxysm! Recommendations


in the right context are key to the success of Spotify here. (Source: http://techcrunch.com/2015/01/12/spotify-now-has-15m-paying-users-60m-overall/ and Spotify app.)

Content—What to Recommend

Recommendations, whether for products, people, or content, make forgreat relevant and personalized content in customer communications.In fact, relevancy trumps the creative quality of the content. In testscomparing the performance of super-creative and meticulously designedemails against the performance of more basic, dynamically created emailswith personalized recommendations we found that the click-throughrates of more relevant emails are three to four times greater than theclick-through rates of the more beautiful emails. Clearly customers pre-ferred relevancy over design—although there is nothing to say you can’thave both.

Increasingly, customers are demanding control over the products orcontent that companies are displaying. Perhaps the most famous exampleof recommendations gone wrong is when Target began sending babyand pregnancy-focused marketing mailers to customers they had iden-tified with a high likelihood of being pregnant. In one instance, Targetwas spot-on but the customer had not informed her parents about thepregnancy and thus the recommendation was highly unwelcome andconsidered a violation of her privacy. We predict that over the nextdecade, most marketing interactions with the customer will becometwo-way, where the customer will be able to provide input and gaincontrol over their own data.

This example has practically become synonymous for all that is goodand bad about predictive analytics. In response, retailers are trying to giveconsumers more control over the purchases brands take into accountwhen making recommendations. For instance, you likely don’t wantthat lice shampoo you bought last week to continue generating rec-ommendations going forward. We get back to privacy in Chapter 17 ofthis book.

As a marketer you want a level of control over the products thatwill be recommended as well. You don’t want your predictive algo-rithms to recommend products that are out of stock, or very low priceditems that are strictly accessories or refills to other products. There are

http://techcrunch.com/2015/01/12/spotify-now-has-15m-paying-users-60m-overall/and


http://techcrunch.com/2015/01/12/spotify-now-has-15m-paying-users-60m-overall/and Spotify app



merchandising or exception rules that you should configure before rely-ing on automated algorithms to populate recommendations on yourwebsite or emails.

Recommender systems know when a product is new, whether it’sgaining traction, or being viewed more or less over time, and are able toadjust recommendations accordingly. For example, when Apple releasesa new iPhone, the new model iPhone is viewed less frequently than theold number in absolute numbers because the old phone has been aroundlonger, but because the latest model is new and trending higher, a rec-ommender system will identify it as more relevant when a user searchesfor an iPhone. Recommender systems take these factors into accountand use temporal knowledge to learn and forget, so recommendationsare up-to-date and relevant.

Beyond Recommendations

Recommendations are most frequently associated with website per-sonalization. However, you can deliver recommendations in anychannel: email, mobile, social, web, phone, or via display advertising.Recommendations thus can be a driving force for both inbound andoutbound communications.

At the same time, there are many other ways to personalize anexperience—on the website or other—beyond delivering productrecommendations. In fact, all the concepts discussed in this book so far,from value-based marketing to life cycle marketing, are opportunitiesto personalize customer experiences—on the web and others channels.What if you could greet a high-value VIP customer on your websitewith a special message? Or if you could welcome back a lapsed customerafter a long absence. Think what the barista would say if you walk intoyour local coffee shop after a long absence. Would he say “hey, youare a woman?”—No! Much better would be: “Hi Omer, it has been along time! I’m so glad to see you back! Can I serve you the usual—anonfat, decaf latte?” Just addressing customers by their first name alonecan have a big impact. A software vendor, Do Inbound, found thatsimply addressing people by their first names in the thank you page of awalkthrough video increased the conversion rate from the walkthroughvideo to account sign-up by three times.

CHAPTER 11

Play Seven: LaunchPredictive Programsto Convert MoreCustomers

Now it’s time to put it all together. In these final three chapters ofPart Two we look at a number of strategies and campaigns you can

use to deliver value throughout the customer life cycle. First we look athow you can use predictive analytics in order to convert more prospects,and we look at how to use look-alike targeting, a predictive technique inits own right, in combination with clusters and other customer segmentsto acquire better customers.

Predictive Remarketing Campaigns

Retargeting or remarketing, which are used interchangeably here, enablemarketers to reengage people who have previously expressed interest in abrand, product, or service by interactions like visiting the brand’s websiteor reading one of its emails. Retargeting is usually associated with visitsto your website, and the subsequent reminder to come back is usuallydirected at you via display advertising. Remarketing programs work in

145


shorter timescales, in hours or days, because their primary objective is toincrease conversions within a specific context created by the consumer.

You probably have experienced these advertisements: after you lookat a pair of shoes on Zappos, and after leaving the site, the pair willfollow you around the web. Whether you are on your Facebook pageor browsing another website—you will often see an advertisement withthe specific pair show up in your feed or sidebar.

The trigger of retargeting however does not have to be limited to avisit to your website. It can also be an email you have clicked, a visit tothe store, or a call to the help desk. Similarly, the reminder doesn’t haveto be delivered via display advertising. The reminder can equally comevia email or phone call. When channels other than display advertisingare used, this marketing technique is usually referred to as remarketing.It seems that some channels are more effective than others when it comesto remarketing. A November 2014 survey of 3,000 consumers in theUnited Kingdom and the United States by research agency Conluminodiscovered that 66 percent of American consumers appreciate an emailoffer related to something they looked at online previously, but only24 percent appreciate receiving this same offer in the form of an onlineadvertisement. When using remarketing strategies, every marketerneeds to consider privacy and the “creepy” factor. This is a very delicatebalance that we cover in more detail in Chapter 17.

When customers visit your website, they directly and indirectly sharea lot of information about their interests and intents with you. You cananalyze not only the number, time, duration, and frequency of their vis-its, but also look at the search terms used to find your site, the specificpages of your site visited, the items looked at, any on-site searches con-ducted, and potentially items abandoned in a shopping basket on yoursite. All of these pieces of information can be used to do personalizedoutreach to these visitors in follow-up communications to try to bringthem back to your company. A simple reminder can often help bringcustomers back to your site.

If the person you are retargeting has been to your website before,and perhaps has even bought from you before, you have even moreinformation to personalize the reminder. For return visitors you can cal-culate their likelihood to buy and the clusters that they belong to, amongother things. Armed with this information, you can further customize

Play Seven: Launch Predictive Programs to Convert More Customers 147

the outreach, increasing your chances of bringing that customer back.For example, if the person who abandoned a search session or a shop-ping cart is a high-value shopper with a low likelihood to buy, you mightas well include a discount to try and get that customer to come back.You have little to lose when the likelihood to buy is low and a lot to gainwhen the predicted lifetime value is high.

You can reach consumers with retargeting reminders using displayadvertising, search advertising (Google Remarketing Lists for SearchAds), Facebook advertising (Facebook Custom Audiences), Twitter(Twitter Tailored Audiences), email, direct mail, or phone calls.

You can only retarget consumers if you can recognize them.You could identify a user based on personally identifiable informationsuch as an email address, or a cookie, a small text file that storesinformation on your hard drive and helps advertisers track you as youmove across the web. If all you have is a cookie, then all you can do istarget display advertising to follow this user around the web. However, ifyou can recognize a user’s email address, from current or past browsingsessions, then you have better options. You can now target this customerusing email, Facebook advertising, or, if you can tie the email to aphysical address, even a postcard.

There are some techniques you can use to recognize more anony-mous visitors to your website. For example, you can tag and capture aweb visitor’s email address upon account creation and subsequent login,or tag and capture their email address on any forms that collect emailson the site—the most typical of which are newsletter sign-ups and guestcheckout (customers provide an email there for confirmation purposes).A lot of the web traffic comes from email clicks, so a major boost inidentification can be gotten when the client includes the user identity inthe email links, so that the identity is passed in the URL (either as rawemail address or as some encrypted ID) or puts code on the site to parseand decode the user ID from the link. Basically your goal is to capturethe identity whenever possible and associate it with the cookie, so youcan identify subsequent anonymous sessions by that user. Using thesetechniques and others, some brands have been able to recognize half ofthe visitors to their website as compared to the industry average of only10 percent of known visitors.

Examples of predictive remarketing campaigns follow.


Predictive Abandoned Cart Campaigns

Abandoned cart campaigns consistently rank among those with the high-est ROI. The Baymard Institute tracks cart abandonment rate statisticsand finds that an average of 68 percent of shopping carts are abandonedon average. Considering this high abandonment rate, all online retail-ers should have an effective abandoned cart reminder campaign in place.An abandoned cart email sees an average open rate of about 30%, ascompared to 14 percentage for broadcast emails. The click rate is about8% as compared to 1.5 percentage and the revenue per email sent for anabandoned cart campaign in retail is about $2.50 as compared to $0.05.

A January 2015 AgilOne survey among 132 retail-marketing exec-utives revealed that slightly more than half of retailers have completelyimplemented an abandoned cart campaign. This was up from 39 percentof online sellers in the same survey the year prior.

For most companies and industries, email is the best format foran abandoned cart campaign. For some large-ticket items, a postcardreminder can be effective. When designing your email, don’t make ittoo complicated. Simply remind the potential customers of their aban-doned cart and include a link to easily bring them back to their cart orcheckout page. You want to get them to the checkout page with as fewdistractions as possible.

Experiment with timing. Measure the response rate to find out howlong after the cart is abandoned to send the message. There are differentschools of thought here. Some technology vendors recommend that yousend a reminder as soon as possible. However, in our experience, fasteris not always better. There are several reasons why you might want towait a couple of hours, or preferably a full day, to send an abandoned cartcampaign. First, some consumers find it creepy that you remind themright away. It strengthens the feeling that you are “watching every moveof your customers.” That is exactly what most marketers are doing, butyou may want to be a bit subtle about it. Second, if you have multi-ple channels it is very possible that some customers abandon a cart butsubsequently buy the item using another channel, for example, your callcenter. Unless your call center software is synchronized in real time to thesystem that sends the abandoned cart reminders, you may want to wait acouple of hours. This is especially true if you decide to include a discountin your abandoned cart reminder. Nothing is more disappointing to a


customer than to realize they could have gotten a discount on the itemthey just purchased. Third, if you send a discount offer too quickly youmay train your customers to expect a discount every time. If customerscan get a discount just by placing an item in a cart and waiting for fiveminutes then every customer might do this going forward. That wouldhave disastrous effects on your margins. Finally, some shoppers don’tappreciate to be bombarded with marketing messages. Whereas theymight have bought from you that day, if you contact them one too manytimes, you might scare them away. As always, it is best to test what worksin your situation.

The essence of a predictive abandoned cart campaign, as comparedto a regular abandoned cart campaign, is the use of predictive algorithmsto differentiate the offer you send to customers. For customers who havea very high likelihood to buy you send a simple reminder. However, forcustomers with a very low likelihood to buy you can safely include adiscount. You have little to lose because these visitors would not pur-chase otherwise anyway. You can go one step further and differentiatethe level of discount based on a customer’s historical sensitivity to spe-cific discount levels. Test the impact of different levels of discount or giftsas part of the campaign to determine whether offering a discount canincrease your revenue and purchase rates, or if it simply eats into yourprofit margins.

Predictive Abandoned Search Campaigns

Unlike abandoned cart campaigns, which are somewhat specific toonline commerce, search and browse abandonment campaigns apply inall industries. AgilOne research shows that visitors who use a website’ssearch functions are six times more likely to make a purchase thanvisitors who do not. These visitors are more than just casual browsers.Don’t lose these potential customers to competitors. You can set upan abandoned search campaign to remind these customers to returnto your site, to give you a try or to get in touch. Abandoned browseemails have open rates very similar, or even higher than, abandoned cartemails, typically in the 30 percent range. Click through rates are about8 percent. The difference between abandoned cart and abandonedbrowse tends to be the conversion rate: about 4 percent for aban-doned browse and anywhere from 20 to 60 percent for abandoned


cart campaigns. As a result the revenue per email that you can expectfrom abandoned cart campaigns is about $2.50, but for abandonedbrowse campaigns it is only $0.50. Of course revenue numbers arehighly dependent on your average order value and the numbers quotedare specifically for retail, where the average order value is about $100.

An abandoned on-site search campaign works just like the aban-doned cart campaign. If a website visitor is logged in, an email orFacebook advertisement can be sent to the address on file offering rec-ommendations similar to or complementary to the searched for items.If the visitor is not a logged-in user, cookies can be used to retargetthem with ads for relevant items.

Abandoned search campaigns also work for people who come toyour site using a Google AdWords search. Google AdWords is often oneof the most expensive marketing channels. Once you have paid for apotential customer to click through to your site via an AdWords query,it makes sense to do everything possible to convert him or her into acustomer. The best part about a visitor coming via AdWords is that youalready have a good idea about what she is interested in buying. Similarto the campaigns above, abandoned AdWords searches can be followedup with targeted offers enticing the potential customer to come back tomake a purchase.

The same recommendations we discussed for abandoned cartcampaigns apply also to abandoned search campaigns, with respect tocapturing site visitor information, timing abandonment campaigns, andincluding an offer based on predictive insights. We recommend youlimit yourself to personalized recommendations and a friendly reminderfor customers with a very high likelihood to buy and that you includean offer or discount for those with a very low likelihood to buy. As withabandoned cart campaigns, you may want to wait at least 24 hours beforesending your reminder or offer, to give customers a chance to completetheir purchase first and to avoid training them to wait for discounts.

Predictive Abandoned Browse Campaigns

Even if you do not have on-site search or if a visitor does not use yoursite’s search feature, you still have the potential to collect data through avisitor’s browsing history. AgilOne data shows 96 percent of all websitevisitors leave your site without buying something. Most of these could


be targeted with reminder advertisements or emails. Compare this to thenumber of people who abandon a shopping cart, which is only 8 percentof all website visitors, or the number of visitors who end up buyingsomething, which is on average only 4 percent. In other words, thereare 12 times more people you could target with an abandoned browsecampaign than with an abandoned cart campaign. For some companiesan abandoned browse campaign generates even more revenues than anabandoned cart campaign. Whereas the likelihood to buy is much greaterfor abandoned cart shoppers, the volume of abandoned browse shopperscan, in some cases, more than make up for that.

When an electronics retailer started to experiment with triggeredemails, they first focused on abandoned cart and post-purchase emails.They sent about 3,000 post-purchase recommendations daily to thosewho had bought something and about 4,000 abandoned cart campaignemails. The campaigns were very successful. Abandoned cart campaignshad an open rate of 55 percent and drove an incremental $10,000 inrevenue every day. When this company launched abandoned browsecampaigns they were pleasantly surprised that this campaign was evenmore successful. By comparison, they were able to send 100,000 emailseach day! Surprisingly, the emails had an even higher open rate thanthe abandoned cart campaign, with a whopping 60 percent of recipi-ents opening the offer. This abandoned browse campaign alone drove$40,000 of incremental revenue each day. This is a good example of pre-dictive marketing in action! Full disclosure: the numbers of this examplehave been changed from the original to obfuscate the company, but theratios are based on an actual case study.

Needless to say, the same rules apply to abandoned browse cam-paigns as to abandoned cart and abandoned search. You can make thesecampaigns even more profitable by differentiating your offer based onlikelihood to buy.

Using Look-Alike Targeting

Remarketing only works for known visitors. Remarketing can help youconvert more browsers into buyers and get past buyers to come backand buy again. However, remarketing cannot help you find and acquirenew consumers for your products, services, or content. This is wherelook-alike targeting comes in.


Look-alike targeting is a predictive analytics technique to find peoplewho look just like an initial “seed audience.” For example, if you feeda look-alike targeting system, such as Facebook’s look-alike audiences,a list of your existing customers, it can find you prospects that have thesame characteristics as your existing customers. You can now use this“look-alike audience” to launch an advertising campaign. The principleis not limited to customers. You could feed a look-alike targeting systema seed list of people who have liked your Facebook page, and it willgo out and find people who have a high likelihood to also “like” yourpage. You can also use this principle to find audiences that behave like aspecific subset of your customers. For example, perhaps you export a listof your best, most valuable customers and then advertise only to thoseprospects who “look like” your best customers. Or perhaps you definea cluster, which likes leather clothing, and now look for prospects that“look like” your leather cluster so you can target very specific advertisingfeaturing leather-clad models to this look-alike audience.

Figure 11.1 illustrates the concept, using Facebook as an example.Facebook look-alike targeting is increasingly popular, but Facebook isnot the only network that gives advertisers the ability to use look-alikeaudiences. Many advertisers, including Twitter, Google Display Net-work, and others also offer look-alike audience capabilities.

On Facebook, you start by uploading a specific list of customers toFacebook custom audiences. This can be the list of customers who preferleather products, or perhaps the list of your best customers. Facebook willnow try to match these records to its user database. Matching happens

Figure 11.1 Facebook Look-Alike Targeting


based on email address. The list needs to contain at least 100 recordsthat match to a Facebook account. After matching at least 100 users,Facebook now uses their internal algorithms, which also use predictiveanalytics, to go out and match your segment to other new people in theFacebook database that “look like” your original list.

Look-alike modeling is a powerful tool that enables marketers togo out and target people who have similar traits or behaviors to theirexisting customers or website visitors. Look-alike algorithms typicallyneed to be fed a list of at least 100 or more existing visitors or customersas a “seed.”

Look-alike audiences can be used to support any business objective:targeting people who are similar to sets of customers for fan acquisition,site registration, purchases, and coupon claims, or simply to drive aware-ness of a brand. It can also be used to find audiences who put an item inthe basket of your website but didn’t pay for it.

Before using look-alike targeting, make sure your seed audienceis clean and well selected or else the look-alike targeting algorithmswill not work. Look-alike targeting is only as good as the inputs are.Remember, here, too: garbage in, garbage out. Make sure your seedaudience converts really effectively before you expand the seed audi-ence with look-alike targeting. Go for quality before quantity for yourseed audience. We recommend that you start with a look-alike cam-paign that is based on your best customers. These are customers that havebought from you several times and therefore you are sure that these arequality customers.

Optimizing for Similarity or Reach

Marketers can optimize their look-alike campaigns for “similarity”or “reach.” When optimizing for similarity, marketers are lookingfor impressions with tight accuracy—and presumably better results.You could say, for example, “with 90 percent certainty, I know thatthis person will buy from you.” There will be fewer of these customersthan customers who have, say, a 60 percent chance to buy from you.When optimizing for “reach” the match is hazier and the ROI lower,but you might acquire more customers overall.

On Facebook, you can choose to optimize your audience for “simi-larity” or “reach” automatically, or to customize something in between.


When optimized for similarity, a look-alike audience will include the top1 percent of people in the selected country who are most similar to theseed custom audience. The reach of the new audience will be smaller, butthe match will be more precise. When optimized for reach, a look-alikeaudience will include the top 5 percent of people in the selected coun-try that are similar to the seed custom audience, but with a less precisematch. Instead of using the types (explained earlier), you can manuallyset a ratio value that represents the top x percent of the audience inthe selected country. The ratio value should be between 1 percent and20 percent and should be specified in intervals of 1 percent. A NorthAmerican Beauty Company used product-based clusters to launch spe-cific look-alike advertising campaigns on Facebook. They first uploadeda list of all existing customers who were part of a bath and body cluster.Then they designed the creative for a Facebook advertising campaignspecifically to appeal to this type of bath and body customer. This com-bination of clustering and look-alike targeting turned out to be highlyprofitable. For the North American Beauty Company, these campaignsdelivered between 2 and 10 times return, when comparing revenuesgenerated to the investment made in Facebook ads.

As said, look-alike targeting works on many other advertising net-works, not just Facebook. The mechanisms for selection and purchasingmedia are similar across networks, and new options become availableevery year.

CHAPTER 12

Play Eight: LaunchPredictive Programs toGrow Customer Value

In this chapter we cover examples of programs that can grow cus-tomer value and revenues after the initial purchase. We cover specific

campaigns including post-purchase programs, replenishment and repeatpurchase programs, new product introductions, and customer apprecia-tion programs, and we take a look at loyalty programs and omni-channelmarketing in the age of predictive analytics.

The Secret to Growing Customer Value

The secret to retaining a customer is to start trying to keep the cus-tomer the day you acquire her. As we discussed in Chapter 7, the mostimportant concept in marketing is to “give to get.” This applies bothbefore and after the purchase. The key to creating customer loyalty isto build value for the customer right from the start of the relationship.This means the customer must have a great experience from the veryfirst interaction.

155


The initial transaction is just the beginning of a long relationshipthat needs to be nurtured and developed. Engagement with customersshould not stop when you convert a prospect into a buyer. If you are ableto convert new customers into happy customers, you can expect futureupsell and referral revenues. Expansion revenue—done correctly—can be very high-margin revenue because it’s typically lower cost todeliver. Referral revenue, too, can be very high margin. Acquiring newcustomers through referrals is cheaper, faster, and more effective thanthrough any other means. Referred customers have a shorter sales cycleand a higher conversion rate than other customers.

The concept of growing customer value during the entire customerlife cycle, both before and after the initial purchase, is visualized in theexpanded customer funnel of Figure 12.1. Remember, it’s a lot easier toget more money from a customer who is happy and already paying youthan it is to get money for the first time from noncustomers. At a highlevel, the path to customer value is: engagement ==> investment ==>offer ==> conversion ==> rinse and repeat.

Figure 12.1 The Complete Customer Funnel

Play Eight: Launch Predictive Programs to Grow Customer Value 157

Predictive Post-Purchase Programs

One of the programs you can use to grow customer value from thestart is a predictive post-purchase program. A post-purchase programis a marketing program triggered by a recent customer purchase.Examples of post-purchase programs are customer welcome campaigns,post-purchase recommendations, replenishment campaigns, and repeatpurchase programs. Post-purchase programs are effective since datashows most customers make a follow-on purchase within a short periodafter a purchase when there is a similar need or the brand is still fresh inthe consumer’s mind.

Customer Welcome Campaigns

The simplest form of a post-purchase program is a new customerwelcome campaign. In retail, our research shows that if a customerbuys only once, the chances that they will come back a second time areonly 30 percent on average. However, if you can get that customer tobuy a second time, the odds increase significantly. Seventy percent oftwo-time buyers will come back again. This means that marketers needto act quickly to reengage new customers and to get one-time buyers tobecome two-time buyers. A new customer welcome message—eithervia email or direct mail—is an easy and effective way to build loyaltywith new customers and build on their early excitement to encouragethem to make a second purchase. A welcome email can see open rates of35 percentage or greater and purchase rates greater than 10 percentage.These rates are right in between abandoned cart and abandoned browsecampaigns discussed in the previous chapter.

The new customer welcome campaign should thank the new cus-tomer, welcome them to your brand, and should include a personalizedoffer to encourage a subsequent purchase. Free shipping for a limitedperiod of time or a free gift with the next purchase can work well withcertain segments of customers. Engaging customers in reviews and shar-ing with friends and colleagues are also effective ways to start buildingrelationships.

A more sophisticated version of a new customer welcome caninvolve a series of emails over a period of time. For instance, a companymight send welcome, thank you, or follow-up emails one day, one week,and one month after purchase. The optimal timing and messaging of


each campaign can vary by company and industry, so be sure test whichcombinations provide the best results for your company. Introducingthem to other categories, additional services, or even uses and care forthe product they just bought are all great ways to continue to buildrelationship with the customer.

Predictive analytics can be used to improve the success of your wel-come program. If you can predict the future value of a customer at thetime of their first purchase, you can tailor your welcome campaign withpersonalized offers geared to different segments. For high-potential orhigh-value buyers, it is worth a lot of money to get them to come back,so you need to carefully craft a set of communications and offers and payspecial attention to this segment. These customers expect a higher levelof service, and for most brands it is important to differentiate treatmentof this segment specifically.

Online services companies—like California-based online homecleaning service company Homejoy—excel at providing best-in-classfirst experiences and welcome messaging to increase customer loyaltyrenewal. Homejoy developed a comprehensive welcome program toensure “100% Satisfaction Guaranteed” and drive people who tested theservices to their second booking. The company sends a series of person-alized campaigns at key moments in the customer journey: a traditionalwelcome email right after the first online booking, a text reminder onthe customer’s mobile phone three days before the appointment, a textwith a special service phone number during the appointment, and afeedback email. One month after the appointment, a personalized fol-low up email is sent with the first name of the client and the name of thehouse cleaner with a special offer. By using different channels to reachcustomers, Homejoy goes above and beyond the client’s expectationsand increases the customer’s likelihood to stick to the brand.

Direct mail is also making a comeback for welcome campaigns. LillyPulitzer sends oversized postcards on high-quality matte paper to everynew customer who bought through their e-commerce site. The creativeis personalized using the first name of the customer. Variable printingtechniques are now available to allow for text or images to be changedpiece by piece without slowing down the printing process. In the caseof Lilly, the offer of the welcome campaign is a “present with yournext purchase” to get the first-time customer to make that covetedsecond purchase.


Post-Purchase Recommendations

Not only new customers should get a follow-up message after their pur-chase. The same is true for any purchase, whether made for a first-timeor a returning customer. After each message, send customers a thank younote and use the opportunity to suggest tips to enjoy the new product orservice. You can also present them with relevant offers or product rec-ommendations for future purchases at this time. Post-purchase emails arejust as successful as welcome emails, and much more successful than anykind of blast message you send. Open rates tend to be over 30 percentand the conversion rate is close to 8 percent.

For consumers, the decision whether to include an offer in thepost-purchase message could be based on the value of the customer.You should invest more in getting higher value customers to come back.For business marketing, the post-purchase program should not focus onthe next sell, but rather on providing guidance and advice on how toget the most out of their initial purchases. Remember, the customerwill only choose to buy more from you if the initial purchase delivers itsexpected value. Therefore, until business customers have received valuefrom your solution, focus on getting those customers to value, ratherthan making next sell recommendations.

Replenishment Campaigns and Repeat Purchase Programs

If your company sells products that have an expiration or naturalreplenishment cycle, such as trash bags or printer toner, send a messageto remind your customers to renew or replenish. You may have noticedthat Amazon now has a “buy it again” button and prominently featuresproducts with short product lifetimes on your personalized homepage.Amazon and other retailers also give shoppers discounts for recurringorders and offer subscription services. Replenishment reminders areamong the most powerful marketing programs out there becauseconsumers perceive well-timed reminders as great customer service,and our research shows that purchase rates can be four times that ofany other campaign. Replenishment campaigns are the best performingof any life cycle marketing campaign. This means that open rates arebetween twenty and fifty percent and purchase rates may be as high asthirty percent as well. For many customers, replenishment remindersare a welcome customer service.


Timing really matters when it comes to repeat purchase reminders.There is likely an average repeat purchase time frame for each of theproducts and services you carry. However, to increase repeat purchaseconversion further, you would need to predict the repeat purchase win-dow for every customer. Products like packaged food items may beconsumed at different rates by different customers, but by tracking therate of purchase for each customer, you can predict the best time tosend reminders for different groups of customers. Replenishment timingcan be set in a combination of three ways: manually set by the mer-chant, wisdom of the crowd looking at average replenishment timingacross all customers, or by looking at individual replenishment cyclesfrom the past. Manual and wisdom of the crowd are good approxima-tions but are no substitute for customer level data because customers mayuse your product at different rates. For example a beauty salon buyinga specific shampoo might need to refill every week, whereas a regu-lar consumer only needs a refill for the same item every three months.Therefore, the most successful strategy is to use customer data whenavailable, then resort to wisdom of the crowd, then ultimately manualentry by the merchant, who can make an educated guess.

An international shoe retailer tested replenishment campaigns onits top-selling products. At first glance replenishment fits better withconsumables than shoes, but it turns out that shoes wear out at fairlypredictable times and loyalists can be persuaded to buy from the brandagain. In one such “shoe replenishment campaign” the brand recom-mended shoes to clients who had purchased exactly 18 months ago, andto a random control group. The open rate of the repurchase email was5 points higher for the test group, visits to the website per email sent were15% higher and product views per visit were 19% higher. Replenishmentemail campaigns proved to be very effective for this brand to createupsell opportunities and to reactivate customers with relevant content.The same retailer sent a successful campaign to parents who bought shoesfor their children also, a specific segment with a frequent need for newpairs as they grow up.

New Product Introductions

When companies release new products or features, they often forgetto advertise these new products or capabilities to their existing cus-tomer base. By using customer segmentation and predictive algorithms,


marketers can now predict which of their existing customers would beinterested in more niche products or feature introductions. And nothingcreates a more compelling reason to buy than new product introductions.

The same strategy can also be used to promote leftover merchandiseor broken inventory. Let’s say you have only sizes 6 and 12 of a populardress that has sold out of other sizes. There is no point in sending a pro-motion for this product to your entire database. However, if you couldpinpoint those consumers who are interested in such dresses and wear asize 6 or 12, you have a highly relevant and powerful campaign.

A New York–based publisher wanted to improve the results of itsNew Releases Newsletters with better targeting. The original segmen-tation this client used was based on genre preferences that customersdeclared when they subscribed to the newsletters. The company thencreated dynamic clusters that were based on the actual purchasing andbrowsing activity of the subscribers. The question: which will turn outto be more accurate, what consumers say they are interested in (pref-erences), or what they show they are interested in through their actions(clusters). It turns out that deciding which genre newsletter to send towhom based on clusters had a two-time better open rate, a four-timebetter click-through-rate and a seven time higher click through rate thanthose sent based on the consumers’ self-stated preferences. The conclu-sion: it is equally, if not more, important to observe what people do ascompared to what they say.

Customer Appreciation Campaigns

It is easy to forget your best customers. These best customers canmake up a significant portion of revenue and nearly all of your profits.On average, we have seen 60-plus percent of revenue and 90-plus percentof profits come from about 20 percent of customers. We’ve met someforward-thinking CMO’s who focus exclusively on their best customers.These customers are the lifeblood of the company, and there are manythings you can do to reward your high-value customers. These rewardsdon’t always have to involve money. For instance, you could give high-value customers a preview of your new collection or invite them to yourheadquarters for a tour. Likewise, your CEO could recognize them witha personal email, call, or handwritten note, or you could give VIPs accessto a special customer service phone number or section of your store.


It might also be wise to recognize these customers more formally suchas with a token gift during the holidays. One of our customers had acompany event where they had a famous singer give a concert. They alsoinvited their platinum customers to this event. It was exclusive andprovided their customers an amazing benefit that money cannot buy.They talked about this in public forums and encouraged other customersto become brand loyalists.

Before you can reward valuable customers, you first need todeeply understand them. Not all valuable customers are created equal.Customers can be loyal in at least four different ways:

1. Customers have been doing business with you for a long time.2. Customers spend significantly with your brand.3. Customers buy from many different categories.4. Customers refer or influence a friend or colleague.

Although all of these customers are valuable, each is loyal in a dif-ferent way and would require a different program to engage with them.

Similarly, when a retailer of outdoor goods and apparel wanted tobetter understand its most valuable customers, it used software algorithmsto find different types of valuable shopping patterns that people mighthave missed. It identified four very different high-value personas that allbehaved very differently:

1. Big-big-but-do-not-return: One group of customers made single, largeorders, but never returned to buy more after the initial purchase.This group likes adult outerwear, as well as high-ticket items like tentsand strollers. They buy high-end brands and tend to live in wealthierstates like New York, California, or New Jersey. These customerstend to be acquired through advertisements featured on vendorwebsites during busy shopping times such as the holidays.

2. Reward addict: The second group of valuable customers has a reason-ably large average order value but not as large as the big-big-group.However, this group buys much more frequently, almost once everytwo months like clockwork. These customers are high revenue butlower margin because they use a lot of rewards points, will not hesi-tate to return things, and will frequently buy items during clearance


sales. These customers buy from the website and tend to be male,young (30–40), and live in less-populated states. These customersbuy outerwear, sleeping bags, tents, and climbing gear.

3. Old school: Old-school customers like to buy from stores, come shop-ping on weekends, and will often buy items on clearance. Old-schoolbuyers have an affinity to the footwear category. They like differentbrands than other valuable segments, and for the most part live closeto the company stores. The old-school customers tend to be acquiredfirst in December through pay per click campaigns.

4. Fallen from grace: The last valuable customer segment discovered iscalled “fallen from grace.” This group of customers typically startsout amazing with more than three orders in their first two monthsand a great average order value that is almost double the average forthe company. However, these customers often stop buying after a fewmonths and have a higher propensity to return items.

Once you understand the different types of valuable customers, youcan start to design campaigns to keep these segments engaged. For theold-school customer segment, you may want to invest in a direct mailor even clienteling campaign whereby VIP customers receive a post-card or phone call alerting them to new footwear models they mightlike. For the Big-big-but-do-not-return segment, perhaps you can send anannual Christmas gift that is an add-on to their tent or stroller, such as ahigh-quality water bottle they can take on walks or camping adventures.

Loyalty Programs in the Age of Predictive Analytics

Loyalty programs are designed to turn customers into loyal advocateswho support and recommend your brand. Consumer marketers havelong used loyalty programs to try to move customers along on thecustomer loyalty continuum.

Retail loyalty programs evolved when progressive retailers recog-nized that without a “customer identification tool,” they were unableto recognize individual customers and reward them for desired behavior.While retail loyalty programs have many purposes, the greatest value thatis created for retailers is this ability to identify individual customers andto measure and understand their individual behaviors. This consumer


behavior data far outweighs the currency value of providing consumersthe opportunity to build a reward opportunity by shopping at one par-ticular store. This opportunity is often misunderstood.

Bolstered by the initial success, Mavi—the apparel company featuredin Chapters 1 and 3—also used predictive analytics to bolster its loyaltyprogram. Mavi implemented a loyalty card as early as 2008. Eighty-fivepercent of its revenues now runs through this loyalty program, and morethan 90 percent of all products bought in the store can be traced backto a specific customer. This is very important to being able to see thecomplete customer picture. Mavi decided to use loyalty points to try andincrease the average order value of a customer’s first order. For example,customers spending an average of $100 would receive a message say-ing “come spend $150 and get extra points.” To another group thatnormally spends $300 it would say, “come spend $400 and get extrapoints.” Not only did the program increase the value of the first order, italso drove repeat visits. Customers now had points to spend and wouldcome for a second visit to do this. Often they would spend more moneythan they had points for. And the average number of visits per customerincreased from 1.2 to 2.1 as a result of this program. The program is sosuccessful that half of the company’s discount budget is now allocated tothis card program.

Loyalty programs are making a comeback in ways that would nothave been possible just a few years ago. First, online purchases and behav-ior can be tracked even without consumers signing up for formal loy-alty programs. Zappos now recognizes VIP customers automatically andgreets them accordingly on their website. You don’t even have to wait fora VIP customer to spend a lot with you to recognize this group. Thanksto big data and predictive analytics it is only now possible to identifybehaviors that point to a customer becoming a VIP in the future so youcan treat them with white gloves from day one. To identify shoppers inthe store, without enrolling them into a formal program, retailers havestarted to use e-receipts, newsletter signups or free wifi as ways to captureshopper information and later identify shoppers in the store.

Second, the days of one-size-fits-all loyalty programs are rapidly com-ing to an end. Instead of offering the same incentives to all customers,offers can and should now be customized on a person-by-person basis.


As the Mavi example showed, the best loyalty programs are tailored todrive specific actions from specific customers.

Third, although loyalty programs have historically focused onrewarding purchases of goods or services, marketers are increasinglylooking to reward behaviors that they know will eventually lead to sales.For instance, the rewards program of online flash sale site Gilt awardspoints to customers who are browsing its website, referring friends,or connecting to its Facebook Timeline because it found that morebrowsing almost always leads to more buying. (Source: Gilt.com websiteInsider Program.) In a way, loyalty is both influencing and influencedby customer engagement. More engaged customers certainly meansmore loyal customers.

An example of a creative and modern loyalty program is thePowerUp Rewards™ program of the company GameStop, which sellsnew and pre-owned game consoles and software and is expanding intomobile devices. PowerUp Rewards members receive many perks thataren’t “points,” such as pre-owned promotional and trade programs,exclusive midnight launch deals, consumer electronics specials, and greatprizes to solidify the company’s position as the destination for all thingsgaming. Members are also automatically entered into monthly EpicReward Giveaways™ when they make a transaction with GameStop.Epic Reward Giveaways are unique and exciting once-in-a-lifetimeexperiences that GameStop created based on their strong relationshipswith gaming publishers and other entertainment partners. GameStophosted more than 5,000 PowerUp Rewards members at their secondannual GameStop Expo in Las Vegas. With more than 200,000 squarefeet of video gaming excitement and innovation, customers experiencedthe features of the new Microsoft Xbox One and Sony PlayStation4consoles before they launched, played the hottest new video games,entered to win great prizes, and met with top video game publishersand other celebrity guests.

Just three years since launch, the PowerUp Rewards program had27 million members in the PowerUp Rewards program, approximately7 million of which were paid members. The program’s paid mem-berships may also include a subscription to Game Informer magazine,additional discounts on pre-owned merchandise in stores, and additional


credit on trade-ins of pre-owned products. Game Informer digital editionhas grown to over 3 million subscribers across 15 countries worldwide,making it the largest digital magazine in the world.

The program has been very successful for GameStop. PowerUpRewards members shop with GameStop approximately 5 times moreoften than nonmembers, accounting for 71% of total U.S. purchasesin 2013. Also, the consumer data that GameStop collects throughthe program allows them to make informed strategic decisions aboutanything from real estate selection, marketing programs, and efficientproduct purchasing decisions (Source: GameStop 2013 Annual Report.)

A Word about Omni-Channel Marketing

A growing percentage of shoppers engage with you using multiple chan-nels, and many shoppers migrate from one channel to another over theirlifetime to become multichannel shoppers. At least twenty-eight percentof shoppers who first buy online migrate to also buy in the store overtime, and twenty-two percent of those who start in the store migrateto also buy online. These percentages are likely grossly understated asmost marketers still struggle to identify in-store shoppers. Omni-channelbehavior represents a special challenge and opportunity for marketers.As we discussed in Chapter 3, the challenge is to create truly completecustomer profiles. The opportunity is to use customer data to drivein-store customers online and online shoppers into your stores. Also,when it comes to predicting customer lifetime value, the number ofchannels customers use is always a very important predictive variable,given everything else left constant.

The role of your various channels isn’t always obvious. Let’s lookagain at the example of GameStop. Rated as one of the top 25 retailers(ComScore Data), GameStop.com brings virtual store shelves to morethan 9 million unique visitors to the site each month. Initially it waslooking at its website as a source of direct e-commerce and revenues.However it turned out that more than 60% of store shoppers visit theGameStop web or mobile sites prior to making a purchase inside thestores, and for every $1 of online sales, the web and mobile channelsinfluence ten times that amount in their stores. The company haslaunched other innovative omni-channel experiences as well that drivecustomer loyalty. The GameStop web-in-store service guarantees every


game is in stock by giving customers the choice to order any productonline while in a store and have it shipped directly to them for nocharge. The pick-up@store service allows customers to shop for games,systems, and accessories online and pick them up at their local store(Source: GameStop 2013 Annual Report.)

As long as you have the physical address of a customer in yourdatabase, you can use direct mail, email, web, or social campaigns to alertall your customers, online and offline, to new store openings, in-storeevents, or in-store promotions. A company that used this strategy suc-cessfully is 100% Pure, an organic cosmetics brand founded in 2005 as anonline-only store and has since grown rapidly to 12 stores in three states.Last year the company sold more than 7 million products. This company’smarketing team is small. 100% Pure interacts with customers across mul-tiple channels, but its web sales account for nearly half its total business.

100% Pure used predictive marketing to promote its brand acrosschannels: using customer data from its online store, 100% Pure was ableto analyze every region in the United States to determine where mostof its customers were located. It found that most customers were locatedin California, followed by New York, Florida, and Texas. This helpedthe company decipher where to open its next seven stores.

Once it opened the new stores, it needed a way to drive onlinecustomers into the stores. Using predictive analytics models, it targetedexisting customers with a high propensity to buy with direct mail todrive traffic to its nearest stores. The company saw a revenue lift of 163percent. For one specific store grand opening, the company sent invita-tions to select customers who lived within a 50-mile radius of the storeand that it had identified as having a greater propensity to buy and ahigh lifetime value. It then sent an email reminder to those customerstwo days before the opening. The store saw sales increase seven times itsaverage daily sales.

One of 100% Pure’s greatest success stories with predictive marketingwas a campaign around its best-selling coffee bean eye cream. Becausethe product is a 60-day supply, the company used AgilOne to trigger anemail at 45 days to invite the customer to repurchase the item online orin store. The company saw a 200 percent average sales lift.

CHAPTER 13

Play Nine: LaunchPredictive Programs toRetain More Customers

In this chapter, we will look closer at the definition of loyalty and churn,and cover specific customer retention strategies including customer

appreciation campaigns, proactive and reactive churn management, andcustomer reactivation campaigns.

Understanding Your Retention Rate

The retention rate of your customers is defined as the percentage ofcustomers that you retain during the measurement period. There areat least two ways to measure retention: you can focus on the percent-age of customers you retain or on the percentage of dollars you retain.We recommend you focus on dollar value retention. This overcomes thechallenge of having great retention metrics at a customer count level andstill having an unhealthy business. The way you do this is by understand-ing the retention rate of your customers by value segment they belongto, as described in Chapter 8.

169


The Concept of Negative Churn

Negative churn is the concept of growing revenue from existing cus-tomers at a rate faster than the rate at which other customers stop buyingyour products or services.

When customers stop buying your products and can no longer becounted as customers, it’s easy to think that the lost revenue will have tobe replaced with sales from new customers and just focus on acquiringnew customers. However, that approach doesn’t take the whole pictureinto account. What you should really be looking at is the total value ofall revenue going in and out of your business, rather than the number ofcustomers you are gaining and losing. That’s because certain customerscould be generating more revenue by buying more products and usingmore of your services, boosting total revenue above the amount of lostrevenue from lapsed customers.

To get to the concept of negative churn we have to move fromcounting the number of customers that leave us to counting the num-ber of dollars that leave. You may lose some customers (and the revenuedollars that come with those lost customers), but if you do your job wellyou will gain many expansion or repeat revenue dollars from other, moreloyal customers.

For example, if you lose $100 through customer attrition and nonre-newals, but you gain $150 for the same period or cohort through expan-sion revenue (upsells, cross-sells, higher usage, etc.) you have negative(net) churn of $50. On a customer-basis (but still considering revenue),if out of 100 customers you lose 10 (gross), but you are able to upsell,cross-sell, or drive additional usage from the 90 customers that are stillthere, allowing you to generate more revenue from the 90 than you didthe original 100, you have a negative churn rate.

When a company is in early stages of growth with just a fewmillion dollars in revenue, customer churn is not necessarily a bigconcern. It seems easy to replace customer revenue with new customers.However, as revenue grows, replacing revenue lost to churn often meanstens of millions of dollars that need to come from new customers.The benefit of focusing early on customer retention and achievingnegative churn (when upsell revenue is greater than lost revenue) is thatthe revenue from your existing customers compounds over time just likeretirement savings.

Play Nine: Launch Predictive Programs to Retain More Customers 171

Understanding Your Business Model

Some industries have inherently higher retention rates than others.In retail, retention rates tend to be very low, typically well below 30 per-cent. In automotive, no more than 40 percent of buyers ever purchase thesame car on two successive occasions. In industries where the relationshipbetween the customer and the firm is more complex, such as in businesssoftware or banking, retention rates can be well above 90 percent.It is no surprise then that the retail industry has been an early adopterof predictive marketing, in a quest to increase customer loyalty andcustomer lifetime value.

Figure 13.1 outlines some of the different marketing environmentscreated by different business models. These environments have differ-ent requirements for customer programs, due to their unique natureof their products, supply chain, selling models, and customer decisioncycles. For example repeat purchase and replenishment campaigns aremore important for those products that have a short product lifetimeand/or a short product replacement cycle.

Type ofBusiness

TypicalIndustries # SKUs

SKUReplace-mentCycle

ProductUsageLifetime Margin

Fashion Electronics,apparel

High(1,000+)

Short Medium(1–3 years)

Varied(10%–60%)

Replenishment/Versioned

Food, drug,cosmetics,CPG

Low-Medium(50–2,000)

Medium Short(<6 years)

High(35%–65%)

Considered White goods,auto

Low(10–50)

Medium Very Long(>5 years)

Low-Medium(15%–-35%)

Subscription Subscriptionconsumerservices,homeservices

Very low(<10)

Low Medium High

Basics Books,electricaltools

Medium Low Medium Low

Figure 13.1 Business Models


Not All Churn Is Created Equal

When measuring customer retention, it is important to realize that notall churn is created equal. For example, churn on new customers is alwayshigher than it is for customers who have already proven to be loyal andwho have come back to do business with you more than once. It isimportant to understand these differences. Let’s look at an example. Youare the head of marketing for a hypothetical golf course called GolfGear.Your CEO has read some industry reports and shows you the followingdata: The churn rate of GolfGear is 15 percent as compared to 5 percentfor competitors. She then adds that even though GolfGear’s growth rateis satisfactory, retention rates are dismal.

Although this seems to make sense on the surface, the reality can bemore nuanced. It turns out the retention rates of GolfGear are exactly inline with the industry average. However, because GolfGear is growingso quickly, a relatively larger percentage of the customer population isbrand new. Because churn on new customers tends to be much higherthan on older customers, the aggregate retention rate for GolfGear islower (see Figure 13.2).

You could dig in a little deeper. When you are losing customers, whoare you losing? Not all churn is equally bad. All brands have unprofitablecustomers. Losing an unprofitable customer is not nearly as bad as losingone of your best customers. Let us look at a large retailer. The retailer isexperiencing a decline in the number of active customers and a declinein revenues (see Figure 13.3).

To truly understand the root cause for the decline this retailer decidedto look at the churn by customer value segment (see Figure 13.4).The overall churn rate observed was 22 percent, but it turns out that theretailer is losing lower lifetime value customers. Churn is highest amongthose customers spending less than $1,000 at 34 percent. For customers

GolfGear Competition

TenurePopulation

(%)Churn

Rate (%) TenurePopulation

(%)Churn

Rate (%)

0–4 months 30% 30% 0–4 months 5% 30%5–12 months 25% 10% 5–12 months 10% 10%1–2 years 40% 5% 1–2 years 50% 5%3+ years 5% 2% 3+ years 35% 2%

Figure 13.2 Example of GolfGear Churn Rate Details


Figure 13.3 Declining Customer Count, but Increasing Aov

Figure 13.4 Decline in Low-Value, Not High-Value Customers


spending between $1,000 and $10,000 the churn rate is significantlylower, namely 15 percent. For medium-to-large spenders between$10,000 and $50,000, the customer base is fairly stable with a smalldecline of 1 percent. However, it turns out that this retailer is succeedingin attracting 16 percent more big spenders, who spend more than$50,000 with the company, which is 50 times more than the lowestvalue segment. As a result, this retailer may very well be increasing itsprofitability even though the number of active customers is decliningand its revenue is decreasing.

Value Migration Is Also Churn

As we outlined in Chapter 8, you do not need to actually lose customersto lose money. The most overlooked aspect of churn is value migration.Value migration happens when customers spend less money with you ina given year than they did in the previous year.

Let’s look at the example of a retail bank. This bank was measuringcustomer gains and customer churn year over year and found that bothwere stable over time with a growth rate of 5 percent and a churn rateof 4.1 percent. Yet this bank was experiencing a significant decline inrevenue. It turns out that the average balance held in the bank by its2 million customers was declining at 2 percent each year.

The management of this bank initially pressured the marketing teamto focus on improving customer churn. However, a simple profitabilityanalysis found that the effect of value migration, in the form of decliningbank balances, was much greater than the impact of customer churn.In fact, of the lost revenue due to value, migration was three times largerthan the lost revenue due to account closings. Once they understoodthe reasons behind the decline in revenues, the bank could take stepsto turn the situation around. It focused the team on combatting valuemigration, not just account closings, and created a separate customersegment to monitor the behavior of the declining segment.

Churn Management Programs

Churn management programs can be untargeted, applying equally to allyour customers, or targeted. Churn management can also be reactiveor proactive. Untargeted churn management could be making general


Figure 13.5 Overview of Churn Management Programs

improvements to service or product quality or undertaking a mass adver-tising campaign. Targeted churn programs can be reactive, triggered bya customer canceling their service, for example, and proactive, based onpredictions of customers at risk (see Figure 13.5). Another name forproactive churn management is retention management.

The advantage of reactive churn management is that you only incura cost for customers who are actually churning. The downside is thatyou may be too late. The advantage of proactive churn management isthat you will likely save more customers because you reach customersbefore they have made their final decision to leave you. The downside isthat you may train your customers to always be looking for deals.

Proactive Retention Management

It is much easier, cheaper, and more effective to try and prevent a cus-tomer from leaving than it is to save a customer last-minute or to reac-tivate that customer after she has already stopped shopping with you.

Central Desktop helps people work together in ways they neverimagined possible through its web-based collaboration platform. Usagedata was stored in disparate systems, making it difficult to identify cus-tomers who weren’t fully using the software. Central Desktop’s servicesteam needed early visibility into thousands of accounts at a time so theycould proactively reach out to those at-risk customers.

Mark Fordham, vice president of services, and Katie Gaston,community and operations manager, led the charge in implementingtechnology to uncover customer insights and become increasingly


forward-thinking as a department. First, they pinpointed key indicatorsof accounts at risk of unsubscribing. For example, they uncoveredfive key features that indicate customer “stickiness” and retention.Customers who use at least two of these five key features turned outto have a 40 percent higher retention rate than customers who usenone or only one of these features. Central Desktop now proactivelypromotes these specific product capabilities to all customers from dayone to increase retention. Additionally, Central Desktop is now able tosend highly targeted and personalized emails to the specific users whohaven’t used all capabilities of the company’s collaboration tool.

Ultimately, predictive analytics combined with proactive customeroutreach has proven to be a winning strategy for Central Desktop: itsoverall churn rate went down by 10 percentage points in just more thanone year. Predictive models can indicate which customers and contactshave a low likelihood to make a future purchase. Predictive models canalso flag customers who are abusing the system. Rather than just lettingthese customers go, you can use preventative touch campaigns to checkin with customers at risk. In consumer marketing that could mean send-ing a simple reminder, a relevant or compelling offer—a personalizedrecommendation, a discount, a gift, or an invitation to an upcoming saleor event. An example of a preventative touch campaign could be sendingcontacts with a low likelihood to buy but a high predicted lifetime value(LTV) an offer for 20 percent off their next purchase.

In business marketing, proactive retention management can be assimple as giving the customer a call at the right time to offer help. Let’s sayyou are a vendor of online software. There could be many reasons thatcustomers are disengaging. Perhaps the primary end user has left and thenew team doesn’t understand the value of the software, or doesn’t knowhow to use it. Perhaps your customer is experiencing performance issueswith your software but hasn’t taken steps to call you for help. Perhapsthe customer has a feature request. There are a thousand reasons, manyof which are preventable and solvable, that a customer is dissatisfied.The problem is that customers may not always realize their problems aresolvable, and you may not realize when customers have a problem.

If you use predictive analytics, you can watch for early warningsigns that a customer is unhappy and check in accordingly. Warningsignals could be declining visits to your website or email opens, orsimply declining consumption of your products or services. For softwareproducts, if a customer logs fewer usage sessions than usual that could


be a sign. By looking at the behavior of thousands of customers whohave churned, software algorithms can identify what is common aboutcustomer behavior prior to canceling their service and alert you if newcustomers are exhibiting this concerning behavior. The end result canbe magical. Nothing is more delightful for a customer than to receivea proactive call from a customer service representative when theyare experiencing trouble. In fact, it is not unusual for these unhappycustomers to be so thankful that they turn into lifetime customers onceyou have resolved their issues.

How Much to Spend to Save Customers?

When it comes to churn management, an important question to ask iswhat is the maximum dollar value of the retention incentive you canoffer for the program to be profitable. How much money you can offerdepends on the lifetime value of the at-risk customers, the targetingeffectiveness (can you accurately identify the churn rate risk of a spe-cific, target segment), and the program effectiveness (churn rate in atarget segment).

Let’s look at an example. A recent cellular network failure in a certainregion has led to many complaints at your call center. You suspect this ishaving an impact on churn. You discover that in fact the churn is veryhigh—around 10 percent—among customers who called to complain inthe past month. So about 10 percent of customers who called in Januaryended up churning in February.

These churning customers are very valuable and carry a $250 averagelifetime value. So you want to create a proactive program to lower churnamong complaining customers. You design a program to call these cus-tomers who have recently complained and apologize for poor service.You also offer them a financial incentive to upgrade to a better servicethe next month. If you know the program can reduce churn from 10percent to 5 percent, then you could spend up to $12.50 on the retentionof each customer.

The general formula to calculate this is:

X = Y (%)∗LTVX = the money you can spend to try to retain each customerY = the percentage of customers that you can save with your campaignLTV = the lifetime value of each saved customer


The trick is that in real life you don’t actually know what Y is, butyou can find Y by testing your program at a smaller scale. All you knowwhen you get started is that the maximum improvement in retention is10 percent (the rate of churn) so the maximum amount of money tospend in this particular case is $25.

The right way to develop a proactive churn management programis to first find customers to proactively target. The way to do this isto develop hypotheses on customer variables that are related to churn.Then you design a test program for these at-risk customers. Make sureto split your target segment in two or more groups, including a holdoutgroup that will not receive a retention incentive. Only testing can ulti-mately quantify how much churn reduction will result, or what level ofredemption will happen on incentives. These learnings can be used torefine the program.

Retention and Share of Wallet

It appears that growing the number of products and services thatsomebody buys from you may also increase the retention rate of thatcustomer. The more effort the customer invests to tell you about them-selves, the greater their stake in making the relationship work. This is awell-known strategy for telecommunications providers. The churn rateson customers who are subscribed only to cable television services ismore than double that of customers who buy their television, Internet,and phone service from the same provider, as outlined in a paper titled“Does Service Bundling Reduce Churn?” by Jeffrey Prince and ShaneGreenstein (November 2011).

AgilOne data shows that the same principle applies across otherindustries: retail customers who buy different categories of productsfrom a given brand have higher retention rates than those customerswho just buy a single product type. Therefore, all marketers should aimto increase the number of product types customers buy as a proactiveretention strategy.

Identifying the Root Cause of Churn

There may be several reasons the customers are not getting the bene-fit from the product or service they have bought. Identifying the root


cause of churn is a best practice when it comes to proactive retentionmanagement. Perhaps customers are experiencing product issues: theproduct frustrates them—poor fit, bugs, data loss, slow performance,irritating user interface. If this is the case, you need to solve the prod-uct issues first, or update your merchandising strategy. Perhaps customersare experiencing poor customer service when they call in to get help.Perhaps the customer bought the right product and you are deliveringgood customer service but they do not know how to use it. In businessmarketing this often occurs when there is a change in the guard and thenew users and sponsors are not familiar with your product.

Whatever the root cause, identifying underlying reasons and improv-ing them can significantly increase your retention rates. Through pre-dictive analytics algorithms you can score each customer’s likelihood tochurn, as well as identify and rank order the factors that contribute tothe churn score.

A consumer services company was able to improve their retentionby ten percent through following steps: First, build a model to pre-dict which customers are likely to churn, and score each customer dailyusing this model. Second, inspect the model to understand factors thatdrive the churn score and strategize around actions to take. Third, testthe effectiveness of different customer treatment methods. Some fac-tors that were driving churn were the number of calls the customerhad to make to the call center regarding trouble with their service,equipment, delivery time, and a sensitivity to price. Customers whowere vocal about the issues they were having were then called back,treated courteously, and offered a discount on equipment maintenanceor on service if they extended their contracts further. For the customerswho were sensitive to price, but were longtime, loyal, valuable cus-tomers, the company proactively called them a few weeks before theircontract ended. The call center representatives discussed their currentcontract price versus the market rate and worked with the customerto ensure they would extend the price they were comfortable with.Naturally—the amount of effort and discount spent to retain the cus-tomers was always proportional to how valuable the customer had beenin the past; or if they were a new customer proportional to their predictedlifetime value.

A subscription-based research firm set out to improve retention oftheir customers. They needed to know the following: Can we predict


the probability that a given customer will renew the annual subscription?More over, if a customer is at risk of churning, what are the contribut-ing reasons (e.g., because the customer is not using x, and the firmdidn’t do y)? Lastly, what levers does the firm have to increase retentionfor each customer at risk, and which of these are the most cost effec-tive? The company analyzed the interactions between customers and thefirm’s primary products, such as search activity and the consumption ofonline research by customers, email and phone conversations with thefirm, etc. A major insight was that consistent engagement of the cus-tomer with the content was more important than just raw volume ofcontent consumption. So a customer that consumes content in a fewbig bursts is less likely to renew than a customer who consistently con-sumes even small amounts of content. The firm now has a consistenteffort to engage each customer at least once a month with relevant con-tent. Delivery of relevant content recommendation by automated emailsturned out to be a very cost-effective retention lever. These emails uti-lized automated recommendation engines to push relevant content—forexample: a customer who read specific content would get an email withlinks to related content. Some high-cost levers (such as consultation callswith an expert rep) are very effective retention levers but should only beused for "premium" customers.

Customer Reactivation Campaigns

When a customer leaves you, not all is lost. AgilOne data shows that it ison average 10 times cheaper to reactivate a lapsed customer than it is toacquire a new one. Reactivation programs for lapsed customers thereforeare a low-hanging fruit for marketers looking for new revenue streams.

Reactivation campaigns are for customers who have not purchasedanything for an extended period of time. Typically, a customer is con-sidered lapsed or lost after she has not spent money with you for twelvemonths. For subscription services, you can consider a customer lapsedor lost as soon as they let their subscription expire—whether after onemonth or three years. Lapsed customers can frequently be reactivatedwith the right offer or product recommendation. Since these customershave essentially been written off and are not expected to make any addi-tional purchases, the most successful reactivation offers are typically fairlygenerous—such as a significant discount on a next purchase.


One company we work with noticed it had a large amount of cus-tomers who once loved the brand but hadn’t engaged with the companyfor a while. It wanted to focus on customer loyalty and engagement bybringing enthusiastic customers back to the brand. The company usedits knowledge of the type of products its customers enjoy to send smartproduct recommendations. These campaigns resulted in an eightfoldincrease in monthly revenue.

Reactivating customers builds on the investments you have alreadymade and avoids the costs of trying to get new customers. These originalcustomers are already aware of your brand and are more receptive, soreengaging them can lead to gaining significant incremental revenue.

In many measurements we made, most reactivated customers behavelike a new customer, that is, they almost restart their life cycle. This alsomeans that the early periods after reactivation are when customers aremost vulnerable to lapse again and require special attention.

Reactivation Campaigns in Four Steps

So where do you start with reactivation? First, determine whichcustomers you want to reactivate. Then determine the most receptivecandidates, customize your message to this group, and reengage thesecustomers using different channels.

Determine which customers you want to reactivate. Not all past customersmay be worth bringing back. Marketers need to carefully determinewhich customers were profitable, interested in products they want tosell/grow, and other strategic factors. For example, you may wantto exclude customers who returned more than 5 percent or had morethan 30 percent discounting on their previous orders.

Determine the most receptive candidates. You may only want to approachpast clients who are most likely or ready to respond—especially if yourreactivation campaign is expensive such as in the case of direct mailor targeted display advertising. Even if you are using email, youprobably don’t want to send marketing messages to customers whoare not ready—otherwise you risk seeming overenthusiastic and drivecustomers away.

Reengage customers using different channels. The shopping journey spansacross many distribution channels. There is no reason your marketingmessage should not be omni-channel as well. After customizing a


message using your customer’s past data, try to reach customers at asmany points as possible. A greater number of contacts usually correlateto a higher response rate. So don’t forget to send that email, postcard,or app notification with personalized offers.

Customize marketing message using past data. Now that you knowwhich lapsed customers are ready to respond, take a look at theirpurchase history. What have they responded to in the past? What typeof products do they buy? What kind of brands do they like? Craft yourmessages according to their tastes and needs. By looking at past trends,you may be able to figure out why the customer lapsed and use this asan opportunity to address this reason. Perhaps the customer may giveyou a second chance.

PART 3

How to Becomea True PredictiveMarketing Ninja

183

CHAPTER 14

An Easy-to-UseChecklist of PredictiveMarketing Capabilities

In order to use the predictive marketing techniques we discussed inthis book, you will need to organize your business around the cus-

tomer and acquire both the mind-set and the enabling technologies forpredictive marketing.

Organizational Capabilities forPredictive Marketing

The mind-set shift required for successful predictive marketing includesa shift from campaigns to the customer life cycle, products to customers,siloed to omni-channel approach, and one-size-fits-all approach to con-textual marketing programs.

Most marketing organizations are still run based on a calendar ofmarketing campaigns: an email blast, an event, a direct mail drop, a dis-play advertising campaign, a store opening. As a consumer, you don’tcare about marketing campaigns. You are one person and you have one,continuous, experience—good or bad—with a brand. The marketingcalendar needs to be replaced with the concept of life cycle market-ing that tries to maximize lifetime value of customers over their entire

185

186 How to Become a True Predictive Marketing Ninja

engagement with the brand through a series of personalized, triggeredcampaigns.

Most marketing organizations still operate in silos. You might havean email marketing group, a store marketing team, a web manager, adirect mail manager, and so on—all working with separate vendors andmarketing service providers. In many organizations the direct mail groupdoesn’t know what the email group is doing and the email group mightnot talk to the display group. In the meantime, customers are receivingmixed messages from different channels.

We all have received multiple display ads, emails, and catalogs fromthe same vendor that were completely uncoordinated. You might havereceived competing offers from different channels that belong to thesame organization. The ads may seem irrelevant and the suggestions forproducts had nothing to do with what you have purchased before andwhere you live. To make sure this doesn’t happen, many companies arereorganizing their marketing teams around providing customers witha consistent message. They are creating omni-channel task teams andappointing heads of the omni-channel experience. Many companies alsouse an “integrated campaign planner” for the purpose of cross-functionalcoordination.

Some companies are starting to organize their marketing team bycustomer life cycle with a separate team focusing on new, existing,and lapsed customers, for example. Other companies assign market-ing teams by customer personas, such as a publisher which organizesaround “romance lovers” and “suspense aficionados.” Whatever orga-nization structure you choose, focus on the customer and the customerexperience.

When Shazam, an online entertainment company, became seriousabout customer engagement, it decided to dedicate a team to this stageof the life cycle. Early on, Shazam realized that relevant communica-tion keeps customers engaged, a key metric for a consumer marketingcompany. Shazam focused on understanding the engagement behaviorof each of its users, such as how consumers interacted with its mobileapplication and how their music tastes differed. When Shazam enabledhighly personalized engagement programs throughout its email mar-keting and app experiences, it increased its customer retention by adouble-digit percentage in the first month of a new registered user,which is huge when you calculate the compounding effect on customerlifetime value.

An Easy-to-Use Checklist of Predictive Marketing Capabilities 187

By now, most people are pretty fed up with “batch and blast”campaigns where all customers get the same message at the same time.These sorts of campaigns lead to high opt-out rates and lower the effi-cacy of marketing campaigns. By contrast, precisely targeted campaignscan triple prospect conversion and customer engagement. These typesof campaigns take into account the customer’s preferences, interests,and propensities to buy so marketers can change their messaging basedon the context of each situation.

When moving from batch and blast to more relevant customer expe-riences, think about these three dimensions for each program, whethertriggered or scheduled:

1. Audience: The first and most well understood aspect of segmenta-tion is deciding who is being targeted and why, what is the context?Are you reaching out because the customer has just visited yourwebsite, or purchased something in a store, or are you trying tore-connect with a customer you haven’t heard from in a while.

2. The frequency of the communication: Part of being relevant is alsoknowing the right cadence of communication. The right amount ofcommunication with your customer leads to higher engagement andultimately higher retention. By reducing the frequency of emails tosome customers, we have seen that marketers are able to reduce theirunsubscribe rates by close to 50 percent.

3. The right content/offer: What is the best content for the customerat any point in time—keeping in mind where the customer is intheir journey with you? In some cases, it could be an offer to enticesomeone to buy something she previously looked at, a reminder tocome back or an interesting article related to something the customeris interested in. For instance, a sporting goods retailer now sends outsnow reports based on the customer’s location to people whom itknows like to buy ski clothes.

Technical Capabilities for Predictive Marketing

Beyond the organization aligned with predictive marketing, you alsoneed to evaluate your technology stack and make sure you have technicalcapabilities to support predictive marketing. In this chapter, we will lookat some of the technical requirements, and in Chapter 15 we will lookat some of the options to satisfy these requirements.


The technical capabilities for predictive marketing include the abil-ity to (a) aggregate and integrate customer data, (b) analyze and predictcustomer needs, and (c) design and execute customer experiences acrossall customer touch points. We will discuss each of these capabilities ingreater detail. Figure 14.1 summarizes these capabilities.

Customer Data Integration

Data integration:

◽ Out-of-the-box connectors for ESPs◽ Webtag for capturing key web events◽ Standard integration packages (ERP, etc.)◽ Turnkey Google analytics integration◽ Enterprise grade integration framework

Data quality:

◽ Fuzzy Phonetic matching (partial match)◽ Standardization (phone, email, address)◽ Daily de-duping (individual & household)◽ USPS DPV (delivery point verification)/

NCOA (national change of address)◽ Genderization◽ Geo tagging

Predictive Intelligence

Out-of-the-box predictive models:

◽ Likelihood-to-buy models◽ Likelihood-to-engage models◽ Clusters (brand, product, behavior)◽ Predicted customer value◽ Product recommendations

Built-in reporting:

◽ Turnkey business and marketing reports◽ Turnkey, configurable dashboard◽ Profitability measures (returns, discounts)◽ Revenue opportunity reports

(benchmarks)◽ Option to connect data to Tableau, Excel

or SQL

Campaign Automation

Campaign execution:

◽ Native e-mail execution (or via partners)◽ Native web personalization◽ Social campaigns (Facebook, Twitter,

etc.)◽ Offline (direct mail, call center, store)◽ Mobile messaging◽ Online advertising (display ad targeting)◽ Web retargeting

Campaign design:

◽ Segmentation and campaign design UI◽ Built-in audiences◽ Built in e-mail creative templates◽ A/B testing (measure results based on $)◽ Measurable hold-out group performance◽ List export (manual and via API)◽ Attribution

Figure 14.1 Checklist of Predictive Marketing Capabilities


Customer Data Integration

In order to segment and target customers, you will need to collect cus-tomer data and integrate it into a single customer profile. In Chapter 3,we described the types of customer data you might want to collect andexplained that data integration is the process of linking all customer data,including online and offline transactions and engagements, into a singlecustomer data profile.

Many marketers start with just online or offline customer data. How-ever, there are huge benefits to starting out with a view that integratesboth online and offline customer behavior because half of all shoppersshop both online and offline. There are also many vendors that can helpconnect legacy systems like order management and enterprise resourcemanagement systems.

Connecting and bringing data together is relatively easy these days;the bigger hurdle is data quality and cleansing. As we discussed inChapter 3, if you don’t link your data you may think you have threeunprofitable customers when in reality these purchases add up to onevery high value customer. Data cleansing is an ongoing effort, andcustomer data ideally needs to be reconciled on a daily basis. The averageAmerican has three email addresses and moves twelve times in their life,so chances are you have a lot of duplications in your database.

One brand we worked with discovered its customer database wasfull of duplications, inflating its customer file threefold. Part of theproblem was that the company had never integrated data from offline,online, and call-center channels. In addition, a long-term promotionit ran that gave discount to first-time buyers had inflated its customerfile since clever consumers signed up with many different names,phones, email addresses. Although it was heartbreaking to see the num-ber of customers decrease, the company now no longer had to wastea lot of money sending three times the amount of promotions to thesame people.

Lastly, do not forget to protect any customer data from theft or acci-dental exposure. Identity theft and credit card breaches are fast-growingcrimes, and unfortunately many companies have made headlines of los-ing customer data in recent years. You do not want to be one of thesecompanies. We will address data protection and privacy in Chapter 17.


Predictive Insights

In Chapter 2 we discussed various predictive analytics models in detail,including propensity models, clustering models, and recommendations.Many technology providers only provide a workbench to create predictivemodels, but some vendors have developed standard models that come outof the box. If models come out of the box, ask what other companiesin your industry have used and tested these models and what were theoutcomes. Also keep in mind that the predictive analytics process is asmuch about data preparation as it is about the development, testing,and deployment of algorithms. Either you have to take care of theseinternally or your service provider or technology provider must do thesethings for you.

Whereas predictive models allow you to divide your customers intodistinct segments, analytics reports will allow you to evaluate the efficacyof the programs you are running and discover new revenue opportunitiesin your existing customer base.

Reports can highlight opportunities, such as customers at risk ofleaving you or a low repeat purchase rate. When a large retailer organizedits data, it was surprised to find out how few customers came back tobuy a second time. In fact, repeat buyers made up only 17 percent ofits customer base. This finding led to discussions at the board level andultimately to a new, companywide, customer relationship managementstrategy to entice first-time buyers to come back.

The more data you have, the more questions you will have as well.Make sure that marketers can access customer data instantly, using aneasy interface, without the help of IT or professional services resources.For example, perhaps you discover you have a very low retention ratefor high-value customers. You might want to perform some root causeanalysis to understand this issue. Do these customers buy from differentchannels or buy different products? Is the return rate on those productshigher than average? These are just some of the questions you mightwant to ask about your business.

Campaign Automation

Ultimately, the rubber hits the road when you, the marketer, design theexperiences that deliver value to customers throughout their life cycle.So ask yourself how you are going to design and orchestrate customer


experiences across channels and across different stages in the customerlife cycle.

Just having insights is not enough. You need to be able to makecustomer insights available at each customer touch point. This meansthat customer insights should trigger customer engagement campaigns,preferably in real time or as real time as possible. Customer insightsshould also be used to personalize dynamic content such as emails ordigital advertisements.

Campaign roadmaps, resource allocation, and project managementshall not be forgotten in the equation for successful execution. Most mar-keters underestimate the bandwidth required to develop and automatemore targeted campaigns. Make sure to define and assign clearly the dif-ferent tasks required and phase the project with realistic timelines.

Questions to Ask Predictive Marketing Vendors

The questions you ask about predictive marketing technology will bedriven by your business requirements. To help you get started, we’velisted some questions that most companies should ask. The questions fallinto the same three broad categories we have just discussed: Will youget a complete and accurate picture of your customers? What kinds ofanalysis will you get, and how can you use it in day-to-day campaigns?

Will I Get a Complete and Accurate Pictureof My Customers?

What channels are supported? Different execution channels providedifferent types of data and require different types of output. Be surethe system can import data from the channels you use today andexpect to use tomorrow and that it can feed data back to those chan-nels in the formats they need. Find out if connectors exist for thespecific channel systems you have in place and, if not, what’s involvedin creating them. If you’ll need real-time interactions, such as help-ing to personalize web pages for individual visitors, ask specificallyhow this is accomplished. Pay particular attention to the support foroffline channels such as direct mail, call centers, or in-store purchases.

What types of data are stored within the system? Every systemstarts with customer profiles. Most can also store transactions.


Systems designed for campaign management will store promotionhistory and responses. Some will capture different types of unstruc-tured data, such as the contents of web pages visited, topics coveredin news articles, sentiments expressed in public comments, andstructured information extracted from such sources. Look for theability to store and reconstruct data that may have changed overtime, such as customer status. Ask whether marketers can add newdata types and sources for themselves. If you sell to businesses, findout whether data is organized at the individual level, company level,or both.

Where does the system get its data? One source is your own sys-tems. These include marketing automation and customer relation-ship management (CRM) and can extend to web analytics, emailsystems, and order processing systems or repositories. Some vendorscapture digital behaviors directly through their own tags on webpages and emails. Several scan public web sites, social networks, andother sources for information that identify companies and individ-uals who are likely to be good prospects for their clients. Vendorsmay also load reference data about companies and individuals fromcompiled directories such as Dun & Bradstreet.

How does the system load its data? Most products offer a combina-tion of direct real-time loads via application programming interface(API) calls and batch loads of files extracted from other systems.Real-time updates are essential if you want treatments in every exe-cution system to reflect behaviors in all other channels. Be sure tofind out if there are any limits to the volume of data that can beloaded, either in terms of response time (how many simultaneousinteractions can the system handle?) or batch volume (will postinglarge files take many hours or incur high costs?).

Does the system provide data quality and enhancement? It is notenough to simply dump customer data into the system. Find out ifthe system can automatically cleanse entries (checking for standardformats, fixing misspelled names, and removing profanity), validateinformation (testing for valid email addresses and mailing addresses,check that the address is not registered as “do not mail” or that theowners have moved, and so on), append likely gender based on firstname, geo-tag based on address, enhance consumer records with


census data or cluster codes, and enhance business records with com-pany size, industry, parent company, and so on.

How does the system link data that relates to the same customer?Linking related records allows you to remove duplicates and to groupmembers of the same consumer household or business. This is essen-tial to building a complete profile and to prevent multiple offers tothe same customer. Linking capabilities vary widely, from sophis-ticated “fuzzy matching” of similar name/address strings to simplyusing identifiers provided by your operational systems. Systems mayalso use external reference data, such as directories of companies andlists of address changes. Ask how much control you will have overmatching and householding rules, but bear in mind that most ven-dors have more sophisticated approaches than users could create forthemselves.

What Kinds of Segmentation and Targeting Will I Get?

What kinds of statistical models can the system apply to cus-tomer data out of the box? Models may predict responses to aspecified offer, recommend what information to present, classify cus-tomers into segments, or serve other goals. Systems vary in the typesof models they build, the amount of human effort required to buildthem, whether the system provides its own model-building tools,and in reporting provided to explain model results. Ask whetherthe models are rules-based (requiring users to define them manu-ally) or statistical (based on truly predictive or automated analysissuch as clustering, collaborative filtering, and propensity models).Rules-based models take much more time to research and configureand are often less accurate than statistical methods. Check whetherthe models are standard or custom and what results other businesseslike yours have seen.

What kinds of analytical reporting and dashboards are avail-able? You’ll want basic customer profiling, promotion analysis,and segmentation. Some systems offer polished dashboards tohighlight trends and current activities. Look for data explorationfeatures such as drill-downs and filters, custom reports in tables andcross-tabulations, data visualization, and trend analysis. Make sure


you understand which data is available to the system’s reportingtools and whether there are any limits on what can be extractedfrom the system for use by other tools.

How Easily Can I Take Actions on the Segmentsor Recommendations?

How does the system help to deliver customer treatments?Some predictive marketing platforms execute marketing treatmentsdirectly, most often by sending emails. But they primarily supportexternal execution platforms by delivering data, scores, or decisions.They may do this on-demand via APIs, by allowing direct queriesfrom external systems, or by sending file extracts. Beyond under-standing capabilities, it’s worth knowing which external systems arealready integrated with a predictive marketing platform and whatfunctions those connectors support.

In Addition to the Questions About Features, Ask YourselfWhether This Is the Right Vendor for You

What’s the underlying technology? Technical information suchas the type of database gives useful hints about likely strengths,weaknesses, and growth potential. You will certainly want to knowwhether the system is offered as a vendor-operated service in thecloud, as on-premise installed software, or both. Also ask aboutthe scale and nature of existing deployments so you can judgewhether your business is likely to make demands the vendor has notpreviously met.

What’s involved in installing and running the system, and howlong does it take? By definition, predictive marketing platforms aredesigned for nontechnical users. But it is still important to under-stand how the system is set up, how long the initial deployment islikely to take, what is expected of the client and what the vendorwill do for you, and what kinds of training and support are available.Judge the skill level and time commitment you will need to oper-ate the system on a day-to-day basis and to make occasional changessuch as adding a new data source.


What help is provided to analyze data, segment customers, orrun campaigns? The services and resources included in the sub-scription or license fees vary greatly. Find out which services areavailable for free and which involve additional charges. Does the ven-dor provide predefined campaigns or a playbook of recommendedactions? Are there ongoing training and tune-up sessions? How manyhours or many sessions with a customer success representative areincluded? Is the customer service staff trained as engineers or mar-keters? How much experience do they have running campaigns likethe ones you are planning?

What will the system cost? Prices may be based on data volume, trans-actions, the number of customers monitored, user count, or otherdimensions. There may also be separate implementation, training,and support fees. Get a detailed quote and be sure it is all-inclusive.Look at whether you will need to sign a long-term contract, whetherpricing is related to performance, and what happens if service levelguarantees are not met.

Who am I doing business with? The background of a system’s devel-oper often gives hints about its suitability for particular purposes,degree of sophistication, scalability, growth path, and likelihood oflong-term survival. Information about funding, number of clients,and time on the market also addresses these topics.

Are you a consumer marketer or a business marketer? Most solu-tions are heavily specialized in either consumer marketing or businessmarketing, which will influence built-in templates, connectors, andeven predictive models.

What is your business goal? Some solutions that claim to deliver pre-dictive marketing are focused on helping you find new prospects oraudiences for your products, whereas others focus more on optimiz-ing lifetime value of your existing customers. Whatever your businessgoals are, make sure the solution you are considering supports them.

Do I have requirements for on-premise software or can I deploya cloud solution? In some companies, security policies or otherrequirements dictate that only on-premise software can be deployed.On the other end of the spectrum, more and more companies findthat choosing cloud-based solutions gives them more flexibility, fastertime to deployment, and lower total cost of ownership.


Do you go with a turnkey or a customized solution? Somesolutions are highly customizable and customized for each andevery customer, whereas other solutions come with many default,out-of-the-box models, audiences, and campaigns.

Do you have true omni-channel requirements? Before you go toofar, decide which channels are your priorities: advertising, social,mobile, web, email, direct mail, call centers, or in-store clienteling.All solutions will have strengths and weaknesses.

CHAPTER 15

An Overview ofPredictive (and Related)Marketing Technology

You have three options if you want to build the technical capabilitiesoutlined in Chapter 14: (1) build predictive models yourself using a

predictive analytics workbench and somehow import these models intoyour campaign management tools, (2) outsource predictive analytics-powered campaigns to a marketing service provider, or (3) evaluate andbuy a predictive marketing solution, such as a predictive marketing cloudor a multichannel campaign management tool. We will look at the prosand cons of each of these three options and also discuss related technolo-gies, which may claim some predictive analytics capabilities.

Do-It-Yourself Predictive Marketing

Predictive marketing technologies have existed for many years in theform of modeling tools such as SAS, SPSS, and Matlab. Many large com-panies, ranging from Netflix, Amazon, or Best Buy, to many companiesin travel and telecommunications have teams of data scientists who usethese workbench-type products to develop predictive algorithms.

However, utilizing these tools has several important drawbacks:before you can use a predictive analytics workbench tool, you first need

197


to translate your business requirements into technical requirementsthat data scientists could turn into algorithms. Most marketers are illequipped to do this, and may not be able to accomplish this task withoutthe help of external consultants. In some ways, you have to reinvent thewheel. From scratch, you have to define what type of models you needfor which business problems, and—unless you are using an experienceddata scientist—you don’t get to benefit from the lessons learned byother companies in your industry.

Utilizing the workbench-type products requires a team of data scien-tists to collect and integrate data, prepare data, develop, test, and deploymodels, and ongoing involvement from IT and data scientists to helpmarketers run reports, extract segments, and prepare lists for campaigns.Many companies have trouble automating predictive marketing usingthese modeling tools.

Most analytical efforts would yield great result if it could be madepart of the workflow of a company. When Omer worked at McKinsey,he helped to create some powerful analysis and algorithms, but as soon asMcKinsey walked out of the room, there was no way for the customer torepeat the analysis or provide the predictive scoring on an ongoing basis.One of the projects Omer was involved in was at Micro Warehouse, acompany for which he later headed up marketing. He created powerfulmodels for optimizing marketing spend and another model for estimat-ing the upside potential of business customers. These models yieldedextremely good results when tested during the project, but when theproject was over, the company could not institutionalize these predic-tive marketing tools. This was one of the reasons the late Jerry York,who was the CEO of Micro Warehouse, recruited Omer to head upanalytical marketing.

Outsourcing to Marketing Service Providers

Database marketing service providers (MSPs) provide a fully outsourcedoption for companies to outsource and analyze their customer database.Most MSPs evolved from providing data hygiene and processing fordirect mail campaigns to offer full hosting and managing of customerdatabases as well as layering on additional services such as analytics andconsulting. For the most part, MSPs focus on servicing larger companies.

Most MSPs will have predictive analytics capabilities, but thesecapabilities tend to be manifested in highly customized models that are

An Overview of Predictive (and Related) Marketing Technology 199

built or tweaked on a per-customer basis by professional services teams.Therefore MSPs are a good option if you have the time and money fora longer-term engagement.

By using an MSP you can reduce your reliance on internal IT. Thereare, however, inherent disadvantages about outsourcing your customerdatabase. You often have to pay per list or per campaign, and MSPs arenot set up to provide customer insights that extend beyond a specificcampaign. Omer has experienced firsthand that every report you wantto run needs to be paid for and scheduled in advance, which often meanswaiting weeks or months for the answer to a question. If you keep cus-tomer analytics at arms’ length in this way, it is unlikely that you willbuild a truly customer-centric organization. For those who are commit-ted to build their company and strategy around the customer, our strongbelief is that it is imperative to bring customer data in-house and to giveall customer-facing personnel instant access to customer data.

Most marketing service providers are essentially consulting organi-zations. Most have the capabilities to deliver direct mail and email cam-paigns, but lack the product capabilities to deliver real-time, personalizedexperiences, such as real-time web recommendations or personalizedadvertising campaigns.

Campaign Management and MarketingCloud Options

Campaign management products are focused on designing and executingmarketing campaigns across channels, including email campaigns, webcampaigns, social campaigns, and mobile messaging. Almost none of thecampaign management tools execute all of these channels natively butwill rely on third parties for some. Campaign management tools certainlycan help you plan the right message to the right customer at the righttime, via the right channel. Many of these tools started as email serviceproviders and therefore have a strong focus around the email channel,rather than the customer. Customer profile capabilities are still very rudi-mentary in many of these systems, though all are certainly focused toimprove these.

Most campaign management tools have a strong online focus, andquite a few have not yet built out the data hygiene capabilities totrack and correct physical address information. Also, not all campaignmanagement tools have robust data management capabilities such as


online/offline identity resolution, fuzzy matching, deduping capabili-ties. Therefore, most campaign management tools cannot be used as asingle source of marketing execution needs.

Most campaign management solutions are evolving to refer to them-selves as marketing clouds, sometimes including some data managementplatform or content management and collaboration capabilities as well.Please note that at the time of this writing, the vendor landscape is inhigh flux. Many established software vendors are offerings a compila-tion of multiple products under one brand umbrella, not well integrated.There are many startups early- and late-stage funded by venture capi-tal firms that are working to build solutions from ground up with thecustomer data being the key element for the connected omni-channelcustomer and marketer of today.

Business marketing–oriented campaign management tools are oftenreferred to as marketing automation tools. Marketing automation ven-dors, too, typically started as email execution for business marketing butquickly added the ability to design web forms and track user engage-ment across web and email. Some marketing automation vendors cannow also design and track social campaigns as well. Marketing automa-tion tools can link customer data from different sources, but tend tolack robust data management capabilities and require matching to busi-ness information from various standardized data sources. Most marketingautomation vendors have some built-in lead scoring, based on rules,but many are not using truly predictive models yet. In business mar-keting, predictive technologies tend to create less value, as the size of thecustomer base is smaller. Like campaign management platforms, market-ing automation tools were developed from a perspective of campaigns,and the architecture therefore tends to be campaign-focused, rather thancustomer-focused.

Other Tools You May Have Heard About

There are several other marketing tools you may be operating or thatyou may have heard about that tout predictive or analytics capabilities.Most of these other technologies are not sufficient for predictive market-ing, but can still play an important role in marketers’ toolkits. For the sakeof completion, we summarize these related technologies in Figure 15.1,which you can use as a reference to lookup “what’s what.”

Typ

eD

escr

iption

Val

ue

Cust

om

erdat

aPre

dic

tive

insi

ghts

Cam

pai

gn

auto

mat

ion

Mar

keting

Ser

vice

Pro

vider

(MSP)

Full

outs

ourc

edm

arke

ting

anal

ytic

san

dda

taba

sem

anag

emen

t

Opt

imiz

atio

nof

dire

ctm

aila

ndem

ail

cam

paig

ns

Cus

tom

erda

taba

seis

outs

ourc

ed,n

otac

cess

ible

toth

em

arke

ter

Adv

ance

dse

gmen

tatio

nco

ntro

lled

byth

eou

tsid

ese

rvic

epr

ovid

er

Dir

ect

mai

land

emai

lca

mpa

igns

;no

real

-tim

ew

ebor

adve

rtisi

ngC

ampai

gn

Man

age-

men

t/M

arke

ting

Auto

mat

ion

Des

ign,

orch

estr

ate

and

exec

ute

omni

-ch

anne

lmar

ketin

gca

mpa

igns

Dem

and

gene

ratio

nan

dlo

yalty

mar

ketin

g;pr

imar

ilyfo

ron

line

chan

nels

Abi

lity

toco

llect

,but

limite

dab

ility

tode

dupl

icat

eor

clea

nse;

exac

tm

atch

only

Lim

ited

pred

ictiv

em

odel

s;of

ten

cust

omm

odel

sre

quir

ing

serv

ices

Dir

ect

orpa

rtne

rba

sed

abili

tyto

exec

ute

acro

ssch

anne

ls;on

line

and

offlin

ech

anne

ls

Web

Anal

ytic

sM

easu

rebo

thag

greg

ate

and

user

leve

ldat

aon

owne

dw

ebsit

epr

oper

ties

Web

site

and

mob

ilepe

rfor

man

ceop

timiz

atio

n

Web

site

met

rics

,can

betie

dto

ape

rson

bym

atch

ing

user

ID’s

with

your

data

base

Con

vers

ion

tren

ds,

web

site

usab

ility

and

indi

vidu

alus

erflo

ws

Pers

onal

ized

mes

sage

son

the

web

site,

via

emai

lor

SMS

tosp

ecifi

cse

gmen

tsD

ata

Man

agem

ent

Pla

tform

(DM

P)

Coo

kie-

and

first

-par

tyda

tafo

rto

pof

funn

elca

mpa

igns

such

asre

targ

etin

gan

dlo

okal

ike

targ

etin

g

Und

erst

and

mak

eup

ofw

ebsit

eau

dien

ces;

use

for

pers

onal

izat

ion

and

targ

etin

gof

web

and

adve

rtisi

ng

Com

bini

ngfir

st-p

arty

cook

ieda

taw

ithth

ird-

part

yda

tato

segm

ent

and

anal

yze

onlin

eau

dien

ces

Som

epr

edic

tive

mod

elin

g,sp

ecifi

cally

likel

ihoo

dto

conv

ert

ofon

line

audi

ence

s

Disp

lay

adve

rtisi

ngce

ntri

c,so

me

abili

tyto

pers

onal

ize

onlin

eex

peri

ence

ssp

ecia

llyw

eban

dem

ail

Em

ailSer

vice

Pro

vider

(ESP)

Des

ign

and

send

emai

lca

mpa

igns

Em

aila

utom

atio

nan

dem

aild

eliv

erab

ility

Just

emai

lbeh

avio

r,so

me

data

clea

nsin

gbu

tno

fuzz

ym

atch

ing

Som

ean

alyt

ics

and

repo

rtin

g,bu

tus

ually

limite

dor

nopr

edic

tive

capa

bilit

ies

Em

ailo

nly;

but

mos

tE

SPs

are

evol

ving

into

full

cam

paig

nau

tom

atio

nso

lutio

nsC

ust

om

erR

elat

ionsh

ipM

anag

emen

t(C

RM

)

Man

age

cont

acts

,ac

coun

tsan

dop

port

uniti

es

Funn

elm

anag

emen

t,sa

les

man

agem

ent

Abi

lity

toco

llect

,but

not

tode

-dup

licat

eor

clea

nse

cust

omer

data

Asn

apsh

otof

the

curr

ent

funn

el,l

imite

dhi

stor

ical

view

san

dno

pred

ictiv

evi

ews

Dir

ect

sale

s,ca

llce

nter

cam

paig

ns,c

anpo

wer

exte

rnal

cam

paig

nm

anag

emen

tA

dva

nce

dan

alyt

ics

Prov

ide

data

anal

ysis

and

repo

rtin

gIn

form

atio

n,vi

sibili

tyA

nyda

taca

nbe

colle

cted

and

map

ped

Pred

ictiv

ean

alyt

ics

wor

kben

ch,b

uild

your

own

mod

els—

notu

rnke

ym

odel

s

Non

e,th

isis

anan

alys

isto

ol

Fig

ure

15.1

Pre

dic

tive

Mar

keting

Tec

hnolo

gie

sO

verv

iew

201


Web Analytics

Web analytics is a critical component to any marketer’s toolkit. There aretwo fundamentally different types of web analytics: the first providesreports or analysis of web browsing data on an aggregated basis, whilethe second type performs person-by-person analysis of named (or anony-mous) visitors. Enterprise class web analytics tools provide both aggre-gate and user-level analytics, ad hoc discovery, and data mining now.Web analytics is mostly aimed to help you improve your user interfaceand site performance. These tools do not aim to deliver omni-channelcustomer experiences. Some advanced web analytics tools are starting tointroduce segment discovery and propensity models, but generally theseare highly custom deployments. Web analytics tools do not aim to buildcustomer 360 profiles by linking customer data from sources other thanyour website, nor execute campaigns across channels.

Data Management Platforms (DMPs)

A data management platform will collect first-party data from a com-pany’s owned web properties, layer this data together with third-party,cookie-based data, and provide analysis about the audience segments ofthe website visitors. This analysis can be used to target specific audienceswith display, search, video, and social advertising campaigns. DMPs dealmostly with anonymous, cookie-based information, and the goal is toacquire more new customers through advertising campaigns, rather thanto optimize the lifetime value of existing customers. DMP platformsare definitely developing into central data hubs, integrating data fromdifferent sources. For the time being, data management platforms andpredictive marketing platforms therefore are very complementary, butover time it is likely that these two data platforms will converge, as isoutlined in Figure 15.2.

Email Service Providers (ESPs)

Email service providers help you design, schedule, and deliver your emailcampaigns. Some email service providers also allow you to integratethird-party data sources and use those to segment your email database andto trigger campaigns. For the most part, data integration is cumbersome,doesn’t come out of the box, and requires extensive professional services.


Figure 15.2 Convergence of DMP and Predictive Marketing

Most email service providers are evolving to become fully functionalcampaign management systems, covering more than just email. Severalemail service providers were recently acquired by larger marketing cloudvendors, and few stand-alone email service providers remain.

Customer Relationship Management (CRM)

Customer relationship management systems were envisioned anddesigned as sales performance management or funnel managementsystems. Despite its name, most systems do not yet manage the cus-tomer relationship. Rather, these systems provide a central repository


of customer data, but most don’t include predictive analytics capabil-ities. CRM systems do not aim to provide omni-channel campaignautomation but can integrate with systems that do. Similar to email andmarketing automation, CRM systems can integrate bidirectionally withpredictive marketing clouds. For example, if you use a CRM systemin your call center, then call center interaction data can be importedinto your predictive marketing cloud, linked to a customer’s profile, andincluded in predictive analyses.

Advanced Analytics

Business intelligence (BI) solutions provide analytics, though typicallythis is limited to backward-looking or descriptive analytics. The goal formost BI tools is to give an overview of your business trends overall, not toanalyze individual customers. You might be able to see the average ordervalue of your customer base and the sales by region, but not what is thelifetime value of a single customer or whether this customer buys fromyou in different stores or just one. You can certainly learn a lot about yourcustomers and your business using business intelligence. Most businessintelligence tools are general purpose, but some vendors are emerging tofocus specifically on marketers with built-in reports and dashboards forthis audience. BI tools can integrate, and often clean, data from differentdata sources. BI tools, however, do not include predictive analyses anddo not aim to provide omni-channel campaign automation.

Which Solution Is Right for Me?

Keep your end goal in mind: make your customers happy by deliver-ing valuable, relevant, and meaningful customer experiences in everycontext. This will build the kind of engagement that leads to profitablerelationships with customers. So ask yourself, what capabilities will helpme deliver the most valuable experiences and build the most profitablerelationships, in the shortest period of time?

The more relevant the predictive models, the more relevant the cus-tomer’s experience because you will have more ways to anticipate andserve your customer’s needs. The more accurate and complete the cus-tomer profiles, the more relevant the experiences. If you can link all past


customer actions to a single profile, you can better anticipate and servecustomer needs. The easier and more accessible the technology is to allmarketers, the more likely it will be used in everyday campaigns.

If you are like most marketers, you don’t have an infinite budget, soyou would like this solution to be up and running as soon as possibleand require as little as possible in terms of data integration and ongoingprofessional services. So in an ideal world, you would have access to aturnkey, out-of-the-box solution, with truly predictive capabilities andadvanced data quality that gives you, the marketer, control over yourown customer data.

Whatever path you choose, make sure to build these three funda-mental capabilities discussed throughout this book:

1. The ability to link customer data from different sources, online andoffline, as well as to prepare your data for predictive analysis.

2. The ability to analyze customer data, using turnkey or custom pre-dictive models, to perform advanced analysis and segmentation.

3. The ability to trigger the right action for the right customer at theright time, across your different marketing execution systems.

Then place this system at the center of any marketinginfrastructure—forming the central nervous system, operating system,or brain of your customer operations. Make sure the system updatesautomatically, at least once a day, ideally without manual intervention,to ensure the latest customer data is there.

Whatever You Do—Get Started

If there is only one thing you take away from this book, let it be thatyou should get started as soon as possible and focus on the very basicsand do them well. Choosing the wrong vendor or the wrong campaignis not as bad as waiting. Your competitors are already leveraging pre-dictive marketing today and gaining a significant competitive advantagefrom their early experiments. Remember that many companies are see-ing the lifetime value, retention, and loyalty of their customers increasedramatically using predictive marketing techniques.

Here are three recommendations.


Get Started Small

With a couple of thousand dollars a month and a couple of weeks of inte-gration work, you can begin solving your customer data problem and runyour first marketing campaign. The best way to build a case for predic-tive marketing is to just get started. Given the large returns expected forthis investment, you really cannot afford to wait. Ask yourself how youwould feel if your competitors deployed this type of technology first.How would your customers feel if they are getting personalized treat-ment from your competitors first, but not from you? Also ask yourselfif there are other projects on your plate that can truly give you a higherreturn on investment.

Bring Customer Data in House but Outsourcethe Data Science

We strongly believe that it is not possible to become truly customer-centric without making customer data available to all customer-facingpersonnel in your organization, starting with you—the marketers.Therefore, we strongly recommend against outsourcing your customerdatabase to a third-party provider such as a marketing service provider.It will be too difficult to access data when and how you want it andthe customer insights will reside outside your organization. Bringingcustomer data in-house does not mean that you need to hire expensivedata scientists or technology resources. Easy-to-use, online solutions areavailable that allow you to own and access your customer data at anytime, but use outside vendors to create the advanced statistical models.Data scientists are in high demand, and most marketers do not havethe bandwidth and expertise to hire, provide direction, and retain suchanalytical personnel. The best data scientists that truly make an impactare the ones that have business acumen, which are even harder to find.Data science is a good way to gain insights, but it is hard to make theinformation available at every customer touch-point without extensiveIT projects. Therefore, start with the end in mind and find the mostpractical solution that gets you to making a difference in the way yourcustomers interact with your brand.


Complement Your Existing Infrastructurewith Predictive Marketing

You don’t have to rip and replace your existing infrastructure. You cancertainly get started by complementing your existing infrastructure withrobust data cleansing and predictive capabilities. Start small and expandyour deployment over time. This could include embedding predictivecapabilities into your different marketing channels, such as email, andperhaps replacing your existing specialized tools with a single campaignmanagement platform to coordinate campaigns across all channels.According to a recent AgilOne survey of 132 retail marketing execu-tives, only 17 percent of marketers use a single solution to coordinateomni-channel campaigns. However, of the remaining 83 percent whodo not yet centrally coordinate campaigns, 42 percent plan to developsuch centralized capabilities in the next 12 months.

CHAPTER 16

Career Advice forAspiring PredictiveMarketers

If you are fearful of data and machine learning, you are not alone.You may have a job in marketing because you love creating amaz-

ing experiences for customers and consider yourself a creative person.Perhaps you were never good at math, and all this talk about data andmachine learning makes you feel uncomfortable. You would be surprisedhow little knowledge you need about predictive analytics to make youan expert! The definition of an expert is somebody who knows moreabout a subject than 95 percent of the population. Just reading this bookwill probably place you in that top 5 percent. Plus, while predictive mar-keting uses predictive analytics under the hood, you don’t need to knowpredictive analytics at all to be a predictive marketing expert.

There is a huge career opportunity that comes from being an earlyadopter of new technologies, like predictive analytics, and new businesspractices, like predictive marketing. There are very few marketers outthere who have direct experience using predictive analytics, or practicingpredictive marketing. This means that even a little bit of experience cangreatly differentiate you in the job market and place you ahead of moretenured marketers. Plus, the demand for data-driven marketers will onlycontinue to increase.

Here are some career tips for aspiring predictive marketers.

209


Business Understanding Trumps Math

Hundreds of programs around the world are popping up at universi-ties with degrees in data science or data analysis. Don’t sign up justyet! We believe it is a real misunderstanding that in order to be data-driven and practice predictive marketing you need to be able to crunchnumbers.

Numbers and stats are useless without people who can draw mean-ing from the data and turn it into strategies, products, and campaigns.This process requires a unique combination of the creative, analytical,and interpersonal skills so often siloed into different departments andjob roles. As big data rises, the need increases for big data marketers whocan draw insight and inspiration from the stats and target consumersaccordingly.

It turns out that finding people who know the business, target marketand customer needs well enough to interpret data is much harder thanfinding data scientists to crunch the numbers. Plus, new technologies arebecoming available that hide the complicated math under the hood andpresent data in a way that is easy for marketers to understand and use.

Although you don’t need to learn to crunch numbers, you still needto feel comfortable using and interpreting them. That means you needto overcome any fear of numbers as quickly as possible. Start by usingand learning simple analytics tools such as Google Analytics or even byunderstanding and reading financial statements of companies you knowfrom your everyday life. Popular books like Freakonomics, NurtureShock,or Moneyball might also help you hone your data-driven way of think-ing by applying an analytics approach to economics, education, andbaseball, respectively.

Ask the Right Questions

So if predictive marketing is about interpreting and using data, howmight a marketer get started with that? The most important thing isto demonstrate curiosity and ask the right questions about your businessand customers. Start with a hypothesis. For example, you might hypoth-esize that you are losing customers because a new competitor is stealingmarket share or because customers are dissatisfied with your latest prod-uct lineup. Once you have a hypothesis it is much easier to go look for

Career Advice for Aspiring Predictive Marketers 211

the data to support or deny this thesis. Any analytical approach is a toolto solve a problem and not a solution on its own. This is very impor-tant to internalize. Many failed projects around analytics are due to thissearch for the magic bullet that never yields results.

Specifically, ask creative, deep questions about your customers.Increasingly, it is the marketing organization that owns customer dataand customer insights. A recent survey of 132 marketing executivesfound that the marketing department is responsible for customer data in75 percent of companies. Management is starting to look to marketingto inform major strategic decisions for the company. This type ofvisibility in the company can be great for your career.

Recently, the director of customer relationship management at alarge discount retailer discovered that a larger than average percentageof customers bought from it once but never came back. In other words,a large number of their customers were “one and done,” which is a com-mon problem in retail. Increasing repeat buyers became a huge growthopportunity for the company. The board of directors of this publiclytraded company discussed these reports. Ultimately the director receiveda promotion and was asked to lead a worldwide team tasked to increasecustomer engagement and customer lifetime value.

Do not just look at the data—mix it with real-life customer expe-riences. Often the best questions come from real interactions with cus-tomers. Do not just stay in your cubicle; get out in the field and interactwith real customers. There is no substitute for customer face time.

Dominique once worked in Japan for Nippon Telegraph andTelephone (NTT) when it had half a million employees. Everyemployee in the company was asked to spend a couple of weekendsworking in the company store to make sure each employee was tunedin to the customer needs. Similarly, Disney asks all new employees—including executives—to work in theme parks in character costume tounderstand the customer experience up close. If your company doesn’thave an initiative like this, you might start one. It will surely differentiateyou and make you, and your colleagues, better marketers.

Blend the Art and Science of Marketing

In an episode of the television series The Crazy Ones with RobinWilliams, a New York advertising agency hires a data analyst against the


wishes of Williams’s character, Simon. The company has a new clientthat sells cat food, and the data-driven marketing campaign designedby the young data analyst outperforms the marketing idea of veteranSimon. Initially Simon and the data analyst clash, but eventually theycome to a happy place of blending the art and science of marketing.In this case, prime time television is not far off the mark. Successfulmarketers learn to combine the science of numbers with the art ofcreativity. Remember: Your job is to differentiate, delight, and disrupt.

Probably the most important thing to realize is that data sciencewill not replace the need for creative thinkers. Dan Pingree, CMO atMoosejaw, described data-driven marketing as a way to inspire and val-idate the creative process. Using data, you can discover new customerpersonas and marketing strategies and test that your creative ideas areworking.

Netflix and its chief content officer, Ted Sarandos, have been outspo-ken proponents of data-driven programming, which they say was behindthe company’s biggest successes in in-house programming, such as Houseof Cards and Orange Is the New Black. However, at a Sundance panel called“How I Learned to Stop Worrying and Trust the Algorithm,” Sarandosconceded: “It is important to know which data to ignore. In practice,it’s probably a 70–30 mix. Seventy is the data, and 30 is judgment—Butthe 30 needs to be on top.”

Learn from Others

There is still a lot you can learn from traditional marketers. Traditionaldatabase marketers, who were focused on direct mail campaigns, arethe most experienced marketers when it comes to predictive analytics.Because it is expensive to send a postcard or catalog, database mar-keters have long used likelihood to buy models and clusters to focustheir mailers on the highest response segments. Digital marketers anddatabase marketers don’t typically spend much time together, but theyshould! The principles used for many years in database marketing directlyapply to modern data-driven marketing. If you have a current or formerdatabase marketer on your team, take them out for lunch and learn aboutadvanced segmentation from them. If you don’t have a database marketeron your team, perhaps find somebody in your LinkedIn network andmake contact.

Career Advice for Aspiring Predictive Marketers 213

You are not alone in your desire to learn about data-driven marketingand predictive marketing. There is no need for you to reinvent the wheel.You could look up companies in your industry that you admire andcontact peers through LinkedIn. Most will be just as keen as you are toget together to compare notes. Start a formal or informal meet-up groupwith other people interested in the field and get together on a regularbasis to compare notes. You could bring in outside speakers to educateyou and your friends. You can even go a step further and make this alarger gathering in the form of a formal meet-up. Leading a meet-upcan be a great way to increase your visibility in the industry and to addrelevant leadership experience to your resume.

A great source of learning is also technology vendors sellingpredictive marketing software. These vendors are working with manycompanies in your industry and can educate you on best practices andbenchmarks which otherwise might be hard to get a hold of. Softwaresales have changed a lot in recent years. Most vendors invest heavily ineducational content, training, conferences, and even industry dinnersand give you access to all these free resources long before they will evertry to sell you something. You should definitely take advantage of thisopportunity and do not feel shy to reach out to relevant technologycompanies—our company AgilOne included. We would love to talk toyou and help you further your career!

CHAPTER 17

Privacy and theDifference BetweenDelightful and Invasive

This is the time to focus on the customer, and build a valuable brandand personal competitive advantage in the process. However, never

forget that you are dealing with customer data, and inevitably there willbe privacy concerns that arise. Marketers today are struggling to walk theline between privacy and personalization, not realizing that these goalsdo not have to be mutually exclusive.

In general, consumers are willing to share preference informationin exchange for apparent benefits, such as convenience, from using per-sonalized products and services. A research paper titled “Personalizationversus Privacy: An Empirical Examination of the Online Consumer’sDilemma” by the Marshall School of Business found evidence thatonline consumers’ intention to use personalization services (and hencetheir willingness to share information) is positively correlated withfactors that build trust in the vendor offering personalization services.These findings would argue that online vendors that seek to benefitfrom their personalization strategies should not only be mindful of theirconsumers’ privacy concerns but should also uncover ways throughwhich they can build trust. In fact, the relative reputation of onlinevendors is one reason why consumers prefer to use personalization

215


from one vendor while ignoring another, even if the services arevirtually undifferentiated. Two important factors known to build trustare the consumer’s familiarity with the vendor and her past experienceswith them.

Here are some more guidelines for dealing with customer data andengendering trust. First, inform online consumers about what informa-tion will be collected how and why. Second, give online consumers achoice about how their information will be used and to which partiesit will be disclosed. Third, make sure you have adequate mechanisms toprotect online consumer information from theft, accidental loss or unau-thorized use. You don’t want to end up on the front page of the WallStreet Journal after a breach. Adequate protection also requires that thereis an effective authority to enforce and impose sanctions for potentialviolations.

Types of Personal Information

When it comes to personalization, there are different types of customerinformation that can be used, and consumers may feel different aboutone type of information over the other.

Anonymous information. It refers to information gathered about pagevisits without the use of any invasive technologies, typically the standardinformation sent with any web or Internet request. Such informationincludes a machine’s IP address, domain type, browser version and type,operating system, browser language, and local time.

Personally unidentifiable information. It refers to “information that,taken alone, cannot be used to identify or locate an individual.” Informa-tion such as age, date of birth, gender, occupation, education, income,zip code with no address, interest, and hobbies fall into this category.The consumer through radio buttons, menus, or check boxes on a webpage has to explicitly disclose most of this information. In addition tosolicited information, personally unidentifiable information also ofteninvolves the use of sophisticated tracking technologies, for example,cookies, clear gifs, and so on. Such technologies, though not identifyinga customer individually, enable the information collecting entity tosketch an effective customer profile.

Personally identifiable information. It refers to information that can beused to identify or locate an individual. These include email addresses,

Privacy and the Difference Between Delightful and Invasive 217

name, address, phone number, fax number, credit card number, socialsecurity number, and so on. Invariably, such information is almost alwaysgathered explicitly from the customer and is typically collected whenconsumers register with websites or engage in financial transactions.

There is some evidence that there is an important psychological dif-ference between companies you have chosen to do business with usingyour personally identifiable data and companies that you have not pre-viously bought from using any data, even if it is anonymous. If you haveeaten in a restaurant before, you appreciate it if you get recognized whenyou return: if the waiter knows your name and remembers that you likeyour meat medium-rare you will be impressed. However, if a strangeron the street suddenly addresses you by name and asks if you enjoyed thewhiskey you drank last night—that is invasive. The same is true on theInternet. If you have bought from Gucci before, you might appreciate itif they recognize you as a high-value shopper and send you a Christmasgift. However, if Gucci competitor Versace is relentlessly targeting youwith display ads that follow you around the web, you might find thatannoying and creepy.

Get Elastic found that 57 percent of online shoppers are comfort-able sharing information as long as it is for their benefit—and beingunknowingly followed across the Internet by unknown brands is hardlya benefit. At the same time Consumer Reports found that 71 percentof consumers state that they are very concerned about online compa-nies selling or sharing information about them without their permission,causing more and more marketers to make the transition from traditionalthird-party cookie retargeting to first-party data.

Avoid Invasive Situations

Use common sense when considering whether a marketing campaign isdelightful or creepy. Consider the context of the situation. Some actionsthat are delightful in some situations might be downright creepy if youchange the context. For example, if a sales representative greets me inthe store it is fine, but if the same sales representative comes to my houseat 10 pm in the evening, it is downright creepy. It seems that customersare more comfortable still with asynchronous communications, such asemail, than with real-time personalization such as display advertising.We shared in Chapter 11 that 66 percent of U.S. consumers appreciate


an email offer related to something they looked at online previously, butonly 24 percent appreciate to receive this same offer in the form of anonline advertisement.

Customer perceptions differ greatly from customer to customer andfrom segment to segment. In general, data from the Columino surveyof 3,000 consumers shows that consumers in the United States weremuch more likely to expect online retailers to personalize experiencesthan those in the United Kingdom: about half of Americans want toreceive a new customer welcome greeting, versus only 34 percent in theUnited Kingdom. Also, shoppers, aged 18 to 34, part of the “millennial”generation, were more likely to appreciate almost all forms personaliza-tion: 52 percent of millennials expect brands to remember their birthdayas compared to 21 percent of shoppers aged 65 and over. Because ofthese differences, it is very important not to generalize: some customerswill outright demand personalization, whereas others are categoricallyturned off by it. What is important is to develop customer profiles foreach and every customer, and to track in these profiles what type ofpersonalization a customer appreciates.

Give Customers Control

Consumers will want control over their own data, and companies thatgive control over their data will win. Large companies like Google andAmazon are already doing this. Giving customers visibility into some ofthe inputs into the algorithm also makes the whole process of predictiveanalytics much less scary. There are three effective ways of giving con-sumers more visibility and control over the ways you are using their data:

1. Edit data: Provide customers with the ability to determine which dataa company can use for recommendations, so the user can erase specifichistory (such as a purchase they made for a friend’s baby shower) orcompletely turn off certain data sources.

2. Explain why: Explain why a specific recommendation is made, so theuser can understand where the relevance comes from. This simpleapproach usually makes even the mistake seem logical and under-standable to the user.

3. Feedback loop: Provide user the ability to give feedback, whether therecommendation was good or bad. This is also called reinforcementlearning that learns not only by observing user behavior (implicitlearning), but by receiving input (explicit learning).

Privacy and the Difference Between Delightful and Invasive 219

Hard Boundaries and Government Legislation

Beyond sound business practice, there are also hard boundaries of pri-vacy. For example, in some cases you cannot collect and sell informationwithout the customer’s permission. There are different industry rules andlegislations designed to prevent accidental information disclosure or loss.In the European Union, data privacy is governed by the Data Protec-tion Directive. American companies wanting to do business in Europeare required to follow the US-EU Safe Harbor Privacy Principles. Anycompanies processing credit card information must comply with ThePayment Credit Card Industry’s Data Security Standard.

The US-EU Safe Harbor Privacy Principles allow for a streamlined pro-cess for U.S. companies to comply with the EU Directive 95/46/EC onthe protection of personal data. The Safe Harbor Principles are designedto prevent accidental information disclosure or loss. Companies oper-ating in the European Union are not allowed to send personal data tocountries outside the European Economic Area unless there is a guar-antee that it will receive adequate levels of protection. There are alsorequirements for ensuring that appropriate employee training and aneffective dispute mechanism are in place.

The Payment Credit Card Industry’s Data Security Standard, called PCIDSS, was developed by the credit card industry. All merchants that handlecredit card information are required to comply with this standard, but insecurity circles the standard is also widely regarded as a good common-sense benchmark for data protection. Even if you are not subject to PCIDSS, you may look to this framework to get an idea how to best pro-tect customer data. Some of the capabilities that are required with PCIDSS are the requirement to install and maintain a firewall, to encrypttransmission of cardholder data, and to track and monitor all access tonetwork resources and cardholder data. In total, the standard specifies12 requirements for compliance.

In some countries, like the United Kingdom, Spain, and Portugal,consent of the individual is required to collect and process informa-tion, and in the case of sensitive personal information, explicit consentis required. The laws in the United States are somewhat of a patchworkacross states, but generally speaking the United States has the least restric-tions on data privacy. Consult with a legal expert, especially when doingbusiness in different countries, because local laws and regulations changefrequently, and many regions, including the United States and Europe,are considering more stringent data protection laws.


We hope that companies will do what is right so that the governmentwill not step in with legislation that is too invasive. One thing is clear:to bolster consumer trust and confidence, companies should protect andtreat customers’ data like it is their own. Businesses must think of dataprivacy not as a back-office activity for compliance, but as a competitivedifferentiator that improves the customer-experience.

CHAPTER 18

The Future of PredictiveMarketing

Upon entering a Rebecca Minkoff store in New York or SanFrancisco, the very first thing a customer encounters is a wall-

length touch screen that offers free drinks. Guests have the opportunityto order a free water, tea, coffee, or espresso. They’re asked for a phonenumber where they will receive a text as soon as their drink is ready.

That perk is not just out of the goodness of Minkoff’s heart.That phone number serves as a digital signature, which tracks themthroughout the store. That large touch screen also lets them browse thebrand’s catalog and put together outfits. While shoppers don’t know it,the store’s employees are plugged into mobile apps, which keep themapprised of who’s in the store and what data is being input into that giantvideo wall. Then, when customers enter the dressing rooms, thingsget interesting.

All the clothing and accessories in Minkoff’s new stores are out-fitted with RFID tags—radio signal-emitting tags frequently used intheme park access wristbands and in credit cards. The dressing rooms atRebecca Minkoff’s new stores in New York, Los Angeles, San Francisco,and Tokyo are equipped with RFID shields that allow them to identifywhich clothing customers bring in to that specific dressing room.

The dressing rooms themselves have mirrors that double as largetouch screens. A computer automatically builds an inventory, basedon the RFID tags, of the clothing a customer brought in with them.

221


The touch screen lets the customer switch to a series of mood-lightingsetups and, crucially, integrates Minkoff’s e-commerce operation intothe dressing room.

Different sizes and colors of the items a customer brings in are auto-matically showcased up on the screen. If something does not fit, theycan order a different-size version to be added to their online cart forfuture checkout. Each piece of clothing they brought in and didn’t tryon is converted into a lead. The innovation lets retailers like Minkoff seewhich items individual customers are not buying. This lets retailers thensend emails to customers urging them to return and take another lookat the clothes they didn’t end up purchasing.

In this book, we have only scratched the surface of what predic-tive analytics can do for marketers. We predict that there will be moreand more use cases of predictive analytics in marketing. Predictive mod-els will become more easily accessible and available to all marketers, atcompanies large and small, as time goes on. Also, predictive models willincreasingly power real-time, personalized communications with cus-tomers both in the digital and physical world. The store of the futurewill be very different than the store we know today, as we saw in thestory about Rebecca Minkoff. Lastly, we believe the widespread accessi-bility of big data and machine learning will spurn a true culture shift ofrebuilding marketing practices around the customer, rather than aroundproducts or selling channels.

Advanced Predictive Analytics Models

In this book, we choose to focus on the predictive models that can havethe biggest impact on customer value in the shortest period of time andare most widely used by early adopters in predictive marketing. However,there are many other predictive models which advanced marketers coulduse. Following are just a couple of examples so you get an idea of howwidely these techniques can apply.

An engagement propensity model predicts the likelihood of a customerto engage with a brand. Engagement can be defined in various ways,but most of the time it describes events such as opening an email,clicking on an email, or visiting a brand’s website. Using an engagementpropensity model, marketers can determine the right frequency ofemails by throttling the number of emails a person receives based on this

The Future of Predictive Marketing 223

engagement propensity. You could also include a directional indicator tothe engagement model to show whether a customer is becoming moreor less engaged with the brand. Marketers can use this model to targetdown-trending customers with special messages and offers to preventthem from leaving the brand.

A total size of wallet model can predict the maximum possible spendfor each customer. This is often called size of wallet or total addressablemarket (we will use the acronym TAM for Total Addressable Market)and is defined as the total yearly spend by a single customer on the prod-ucts/services sold by a company. A TAM prediction can also be made fora specific product category. Marketers can use these models to identifycustomers who are likely to spend more with the brand. As with engage-ment, you can complement a TAM model with a Direction of TAMmodel. A Direction of TAM model predicts whether the total address-able market for a given customer is growing or shrinking. For example, ahot start-up company may have a low TAM on servers today but its direc-tion of TAM may be very high, indicating it is growing fast and couldbecome an important prospect soon. Marketers can use this model toidentify hot prospects.

A pricing optimization model predicts the price that best drives sales,volume, or profitability. The model should be customized based onwhat you are optimizing for—sales, profits, volume, or any other factor.A marketer can use a pricing optimization model to decide the bestprice for a given product or service for each customer. A differentmodel can be developed for each product you want to do this for,or a more generic model can be used that predicts price sensitivityof a customer in general. Similarly, an offer optimization model candetermine which offer will have the most impact on which customer.This kind of model can be configured to maximize conversion, revenue,or margin. This model helps marketers to send the right offers to theright people.

A keyword-to-contact recommendation model can predict the affinity ofa customer to certain content, such as a newsletter or email, based onthings like web behavior or past purchases. Similarly, a hot topics detec-tion model would predict what are the hot topics or hot products thatcustomers are interested in based on analysis of a customer’s social activ-ity, website logs, sales, and other sources. Marketers could use thesepredictions to decide on content marketing topics for specific customers.


A predictive clustering model predicts which cluster a customer will fallinto in the future. Marketers can use these predictions to start differ-entiating treatment of customers right after they have been acquired.There is no longer a need to wait for them to express their behavior tounderstand who they are. Instead, marketers can react before it is toolate when customers are moving to lower value customer clusters.

Think Like a Predictive Marketer

Something struck us in our work with successful early adopters of predic-tive marketing. These people took a different approach. There is a strongwin-win here for consumers, companies, and individual marketers alike.Customer satisfaction is up, companies have been transformed aroundthe customer, and individual marketers have gained a higher profile intheir companies. One company we worked with was able to doubleoverall revenue by growing its online revenue around a brand that wasrooted in stores. Online purchases rose from 20 percent to 80 percent ofrevenues in the process.

So we finish with some advice, not on what to do, but on how to be.Here are some lessons learned from predictive marketers that will helpyou make this amazing transition.

First, if there is only one lesson you take away from reading thisbook it is to focus on the customer and organize your thinking aroundthe customer. Customer-centricity is not a new concept. In fact, mar-keters have strived to focus on the customer since the dawn of market-ing. Just because the concept has been around for a long time doesn’tmean it has been implemented successfully. And just because your com-pany might have failed to build a true customer-centric culture so fardoesn’t mean you shouldn’t try again. Big data and predictive analyt-ics truly enable a customer-centric organization in a way that wasn’tpreviously possible.

Second, focus on actions, not analysis. Too many analytics projectsfall flat because companies get stuck in the proverbial analysis paralysisphase. Once you get used to numbers, you can get hooked on them.Every revelation about your customers will prompt another question.We only have a low percentage of repeat buyers. Why is that? Are the


defectors young or old? Do we have more defections in the store thanonline? Are there certain products that bring in more repeat customers?Are the defectors returning more products than others? There is noend to the questions you could ask, and answer, using customer data.However, there is a real risk in asking too many questions all at once.The risk is that you do nothing with the data. Just having data alone willnot change anything. Only ask questions that will lead to an action, oth-erwise, it’s all nice to know. This is what a lot of consultants do, pages ofanalysis without any action, all interesting facts. You will only improvecustomer experiences, lifetime revenue, and your company’s financialperformance if you act on customer data. This means using even justone customer insight to change something, to launch one campaignthat might improve things. In this example, perhaps it is to launch orimprove a customer welcome campaign with post-purchase recommen-dations and see if it makes a difference. The sooner you start acting oncustomer data, the sooner data-driven marketing becomes a way of life.Also, it helps to have some early wins, in the form of improved customersatisfaction and revenues, to give you the mandate inside your companyto continue your data explorations.

Third, get started today, keep it simple and iterate. It is true thatthe technology landscape of vendors supporting predictive marketing ischanging rapidly. Newer and better technologies will become availableevery year. However, that is not a good reason to sit and wait. There isa huge benefit in being an early mover. The companies featured in thisbook know that. A major challenge we had in writing this book wasto get marketers to agree to share the results they have achieved withpredictive marketing. Often we heard: “We don’t want our competi-tors to know just yet how big the rewards are from deploying predictivemarketing” or: “We view our predictive marketing initiatives as a majorcompetitive advantage and we do not want others to know about it.”This tells you something. Take a risk and get started with predictivemarketing—even if it is at a small scale. It will take your company a whileto get used to the new way to thinking required for predictive market-ing, so the sooner you can start this transformation and learning process,the better off you are. Eventually, your customers will demand that youserve them more relevant content, and as your competitors are rolling out


predictive marketing, you may start to lose customers to them. If youknow you will have to adopt predictive marketing eventually anyway,why not start today and gain a competitive advantage?

Fourth, frequently try new things and measure everything. Don’texpect that you will crack the nut on predictive marketing on the firstattempt. It will require some trial and error to find the data and thecampaigns that work for your company, but the only way to learn thisis to get started. However, make sure you formulate a clear hypothe-sis for every experiment and test your hypothesis by A/B testing or byalways including a holdout group that doesn’t receive the new treatment.Improve gradually by testing new things. Be patient. It may take a whilefor your company to catch on. Don’t stop after the first failed experi-ment. Some of our most successful clients run hundreds of campaigns,each contributing a small percentage to the overall experience and life-time value of their customers. Like football, predictive marketing is agame of inches. Every marketing play helps to move the ball forwarddown the field, toward the goal of loyal and profitable customers.

Fifth, communicate successes inside and outside your company.As you start to see results with predictive marketing, share these withthe world. There is a real advantage in finding a group of like-mindedmarketers—inside your industry—and comparing notes. Perhaps youcan find marketers at companies or brands that you have a lot incommon with but don’t directly compete. Your learning will accelerateif you don’t have to invent the wheel by yourself. Predictive marketingis turning into a powerful movement, and finding like-minded prac-titioners can be both profitable and fun. We recommend you also tellcustomers about your experiments, perhaps using your blog or by doinga press release. We would be happy to put you in touch with reporterskeen to write about this movement. Customers will like reading aboutthe fact that you are investing in better understanding and servingthem, as well as efforts to eliminate irrelevant mass marketing practices.Lastly, blog or tweet about your successes so you can get the credit youdeserve as a forward thinking, modern marketer. Why not build yourown reputation in the process of adopting predictive marketing?

These are the principles of predictive marketing. We are very luckyto get to witness one of history’s most profound technology and cul-tural revolutions. Aspire to these qualities and you can use the strategieswe’ve laid out to your advantage—or invent your own. You’ll be able to


build on your successes, both with customers and within the company.And then, as predictive marketing rises and becomes ubiquitous, youwill be ready.

Marketing is on the forefront of a much larger revolution. Machinelearning will eventually permeate all walks of life and improve the qualityof education, philanthropy, and health care, among others. As marketers,we are the pioneers who are exploring how to best use the intelligence ofmachines to improve our lives—without invading privacy in undesirableways. What you do truly matters. We wish you the best of luck in yourpredictive marketing endeavors and encourage you to tell us about yourexperiences:

Website: www.predictivemarketingbook.comPredictive Marketing Book LinkedIn group: www.linkedin.com/groups?

gid=8292127Twitter: twitter.com/agilone

Please stay in touch and join the conversation!

http://www.predictivemarketingbook.com

http://www.linkedin.com/groups?gid=8292127

APPENDIX

Overview of CustomerData Types

Purchases and Transactions

You can track online purchases by a tag you install on your websiteor directly from your order management system, which could be youre-commerce platform. You can track in-store purchases from your pointof sale (POS) application or an order management system. Often, storepurchases are collected by a different system than online or phone pur-chases. It is extremely important to be able to tie all purchases back to thesame person. For example, if you send abandoned cart reminders with-out reconciling online and phone purchases, people who abandoned acart online, but completed the purchase over the phone, may receivea discount coupon for checking out their abandoned cart. These cus-tomers are going to be upset and complain: “You mean that if I hadwaited, I could have received a discount for my item?”

Web and Online Behavior

If you are a business marketer, web browsing behavior and email interac-tion may be even more important than purchases. These two behavioraldata points give you a good idea of the future intent of a prospect orcustomer. Purchases alone would not give you enough information inbusiness, given that there are typically much fewer purchases than in

229

230 Overview of Customer Data Types

consumer marketing. On the flipside, the decision cycle in business mar-keting is much longer so there are more of these prepurchase data points.In fact, predicting who will purchase in business marketing is so lucrative,that a cottage industry of specialized “predictive lead scoring” vendorshas sprung up to help marketers know which prospects are getting readyto make a purchase. With this information in hand, companies can focustheir limited sales resources on the customers that matter most.

You may not be able to recognize all visitors during their first visit;however, using cookies—tags placed on a customers browsing activity—you can start tracking the behavior even for anonymous visitors and tiethis back to real people later. You will know who people are when theyidentify themselves, for example, by making a purchase or by signingup for your newsletter, or by filling in a form before downloading anelectronic book or watching a video from your website. You can useprogressive profiling to collect customer information without hurtingconversion rates. Instead of having to fill out 15 form fields in order toregister, a site visitor only visits a couple of fields at a time. For examplethe first time customers come to the site, they fill just first name, lastname, and email and on the second visit are asked for phone numberand zip code, and so forth.

Email Behavior

Email opens and clicks can be an important signal of user engagementand likelihood to buy. You can learn much about customers by ana-lyzing which emails they read and how often they open your mail.This information can even be used to make predictions about the likeli-hood of an individual to unsubscribe from your email list. Unsubscribesare seemingly innocuous, but are actually very expensive. We come backto optimizing email frequency in Chapter 9 of the book.

Household and Account Grouping

Householding was an important data processing step in catalog market-ing, where marketers didn’t want to send the same catalog twice for twobuyers from the same household. Not only does householding help toreduce direct mail costs, it helps you better understand truly valuablecustomers. By myself, I may not look like a very valuable customer, but

Overview of Customer Data Types 231

if I directly influence the purchases of a number of household membersI may very well be a VIP.

The relationship between individuals and the households theybelong to is not always easy to see. Using the person’s name andaddress, likely relationships between family members can be established.Software can even be used to automatically assign a head of household,which is the person deemed most influential in that household when itcomes to buying from your brand.

In business marketing, the relationship between a contact (a prospector a customer) and the company he or she works for is often recorded inthe company’s customer relationship management system. For new leadscoming in, you may need to invest in custom code or third-party toolsto automatically link contacts to the account they belong to.

Location

After customers make their first purchase, their home address can bederived from the billing or shipping address. It is useful to have soft-ware automatically calculate and append the longitude and latitude ofthe location of a customer’s home address. This will allow you to makeother calculations later on and to target customers based on attributeslike their proximity to a certain store. For example, when you have anew store opening you could send all customers within a 10-mile radiusa postcard invitation to come by and check out your new location.

When 100% Pure, a cosmetics retailer that started as an online-onlybrand, decided it wanted to open up physical stores it turned to thedata it collected on existing customers. Based on their shipping andbilling addresses collected when they bought products online, the com-pany chose its new physical store locations. You can read more about the100% Pure experiences with predictive marketing in Chapter 12.

For business marketing, company address is also very important, assalespeople often get assigned based on the location of a customer.

Beyond a customer’s home address, it is also useful to collect moretemporal location information. In other words, it is useful to knowwhere the customer is right now. Web logs can be used to look up acustomer’s IP number and marketers can determine a customer locationbased on this. You might have experienced this technology most vividly


when you travel overseas and Google is asking you if you want to switchthe default language of your search engine.

Mobile technologies such as iBeacon determine a smartphonelocation with high accuracy. Supermarkets such as Marsh and Safewayhave been early adopters of iBeacon technology. Studies conducted byCoca-Cola and Procter & Gamble among others show an immensedifference in effectiveness between offers that are made to consumersright in the store versus at home or elsewhere. The time-of-purchasepower is very important. For example, beacons in Marsh stores will beable to trigger alerts such as shopping lists, ads, and other content forcustomers who use Marsh’s mobile app. And because beacons are moreaccurate than GPS, it can send the alerts when you are in the right aisle.

Call Center Interactions, Meetings, and SocialInteractions

Customers may be making phone calls to your company, providing evenmore important data points. New text analysis software can help yourecord and analyze the length and frequency of calls, as well as the topicsof conversation and customer sentiment. The easiest way to incorpo-rate call center interactions is to classify each call into a category and toappend this category to each customer data profile.

Especially for business marketing, the fact that a prospect or a cus-tomer agrees to meet with you over the phone or in person is a strongbuying signal and you want to record this in your database. As withpurchases and call center interactions, there are a large number of datapoints—direct and derived—that can be recorded about each meeting.Think about the time, location, and length of the meeting as well as thecustomer sentiment during the meeting.

Social interaction poses the same problems as call center data.These are typically natural text conversations, and in order to appendactionable information to a customer profile, you will need to classify ortag the social interaction. Vendors are trying to mine sentiments fromsocial comments, such as whether a customer is making a positive or anegative comment. The problem so far has been that most of these senti-ment analysis packages are highly inaccurate and can give marketers moremisinformation than helpful data points. Large consumer brands often


resort to assigning customer service personnel to monitor social feeds andto manually classify and assign conversations.

Returns, Complaints, and Reviews

Returns and complaints hold very rich information about customers’likely retention and advocacy. Not many customers return or complainand therefore it is sparse data, but it should be highly paid attention to.For example, it turns out that in order to predict customer lifetime valueor customer retention, returns or complaints are among the top five mostimportant variables.

Reviews and surveys can provide information that is equally valuableto complaints and returns. The reason we recommend you don’t startwith analyzing reviews when considering predictive marketing is thatoften reviews and surveys are left at third-party sites. It is trickier tointegrate with these sites and tie back the reviews to a specific customer.Some of this might violate customer privacy, and not all customers arecomfortable with it.

Gender

Segmenting your customers by gender is one of the most basic segmen-tations you can do. It is easy to create a newsletter with a dynamic headerthat will change based on the gender of the recipient. This has provento significantly increase click and conversion rates. Also, there really isn’ta point to send feminine products promotions to men and shaving pro-motions to women. Typically the gender of customers isn’t collectedspecifically as part of a purchase but software algorithms can automat-ically recognize most names and tag many of your customers as menor women.

Beyond targeting men and women with different marketingand offers, men and women also have different shopping behaviors.We recently analyzed 1 million consumers visiting gift and variety sitesand found the following differences:

• Men have a 24 percent higher lifetime value than women becausethey shop more often and have larger transactions.

• Men are twice as likely to buy using rewards points than women.


• Men are slightly more expensive to service, with margins for menbeing 4 percent lower than women, because they use more rewardsand more discounts.

• Men are slightly more likely to buy across multiple product categories,such as socks, pants, and watches rather than buying only pants.

• Men are slightly more likely to buy from Amazon rather than from abrand or specialized retailer’s site directly.

If you are a marketer for one of these sites, you may want to offerrewards points to male customers, as they are very receptive to rewardsprograms. You may also want to focus your retention budget on femalecustomers who are more apt to return and buy frequently.

U.S. Census Data

U.S. census information offers an important source of data enrichmentthat is often overlooked. U.S. census data is freely available to everybodyand can be matched with your customer records based on zip code.

If you know how many people live in a certain area you can com-pare that to the number of customers you have in that area. So now youcan essentially calculate your market penetration for a specific region.Based on this information you may decide to increase your acquisi-tion budget for regions with low penetration. You can learn from U.S.census data what kind of housing is most popular in a specific neigh-borhood. Especially if you are marketing lawn mowers, it could be veryimportant to know whether this is a neighborhood of apartments orsingle-family homes. You can approximate household income based ona customer’s zip code. Household income has always proven importantfor understanding customer behavior.

Vertical and Size

In business marketing, the size and industry of a company is probably themost frequently used demographic data for segmentation after location.Third-party databases can help to augment your data with the right com-pany size, vertical, and number of employees. Traditionally companies


like Dunn and Bradstreet and Harte Hanks provided this data. Increas-ingly, the most accurate and up-to-date view of company size, at leastthe number of employees, is LinkedIn. LinkedIn will not have revenuedata, but probably has an up-to-date count of the number of employees.Industry vertical and employee size in business marketing become veryimportant when calculating the total available market or the share ofwallet or market share. Business marketers typically don’t care about theoverall penetration of a market, but about the penetration of a specificmarket segment—by vertical or company size.

Other Customer Data Points

The amount of data you can collect about customers really has no end.For example, there are many third-party data sources that could betapped to enrich your customer data. One popular example is to mashup a customer’s location with the weather that is predicted for that area.If you could collect both in real time, you might surface umbrellas tothe front page of your website just like you would in a physical storeduring rainy periods. One retailer we work with has experimentedwith weather-based campaigns, but has not yet found a profitable wayto leverage it on its site. First of all, it is not trivial to create campaignson the fly in response to weather, and second, it has not yet been foundthat this increases sales materially.

INDEX

NOTE: Page references in italics refer to figures.

Abandoned cart programs, 24,148–149

Accenture, 54Account grouping, 47Acquisition

managing customers, 70, 70–71marketing spending optimization

for, 78–86, 79, 80, 81, 82,83, 84, 85

using clusters to improve, 99–100value-based marketing for,

115–116, 116Active customers, strategies for,

112–113Addresses

collecting data, 231–232validating data, 52–53

Address type flag, 52Advanced analytics, 204Advertising spending, conversion rate

and, 81, 81–83, 82, 83AgilOne

on abandoned cart/searchcampaigns, 148, 149

on adoption of predictivemarketing, 14, 16, 18

on growing customer value, 156inception of, 19on marketing spending

optimization, 74

Alain Afflelou, 139Amatriain, Xavier, 4Amazon, 4, 33–34, 113Anonymous website visitors,

recognizing, 147, 216Apple, 111, 144Application programming interface

(API), 192Appreciation campaigns, 161–163Arcelik, 94, 125Artun, Omer, 19Audience, understanding, 187Average Revenue Per User

(ARPU), 72

Baymard Institute, 148Behavior of customers

behavior-based clusters, 94, 97–99,98

buying behavior and outlierdetection, 37

email behavior and datacollection a, 47

likelihood to engage modelsand, 130–136, 132, 133, 134,136

Benchmarking, for marketingspending, 80, 81

Best Buy, 19‘‘Bias-variance dilemma”, 39

237

238 Index

Big-big-but-do-not-returnsegment (high-valuecustomers), 162, 163

Big data, 3–21Birge, Robert, 139Birthdays, of customers, 38Blattberg, Robert C., 120Bosch, 43–44Brand management

brand-based clusters, 94–97, 97customer relationships and, 14–20

Browsing, abandoned, 150–151Burlington Coat Factory, 106Business models, retention

and, 171, 171Business understanding, need for, 210

Call centers, data collectionand, 47

Campaign automation, 190–191,199–200. See also Informationtechnology (IT)

Carts, abandoned, 148–149CASS certification, 52Central Desktop, 175–176Cetiner, Bora, 94Channels

finding channels that bringhigh-value customers, 88–89

last-touch attribution, 89–92, 91for making recommendations to

customers, 144omni-channel marketing, 166–167questions for data collection,

60–61vendor support for, 191

Checklist of predictive marketingcapabilities, 185–196

organizational capabilities,185–187

overview, 188

questions to ask of vendors,191–196

technical capabilities for predictivemarketing, 187–191

Churnchurn management programs,

174–175, 175differences in, 172, 172–174, 173negative churn concept, 170preventing, 74See also Retention

Classifier and system design, 39–40Clienteling, 49–53Cloud options, 199–200Clustering

avoiding mistakes and, 100cluster DNA, 94defined, 23, 25to improve customer acquisition,

99–100insight from, 101models, 25–28, 27overview, 93–94predictive clustering model, 224segmentation compared

to, 26–28types of, 94–99, 96, 97, 98

CMO Club, 54Collaborative filteringrecommendations, overview, 33–34types of recommendation models,

34–36, 35Complaints, data collection and, 47Conlumino, 146, 218Content

determining, 187making recommendations,

143–144Context, customers and, 141–142Conversion, 145–154

abandoned browse campaigns,150–151

Index 239

abandoned cart campaigns,148–149

abandoned search campaigns,149–150

conversion rate and marketingspending, 81, 81–83, 82, 83

look-alike targeting and, 151–154,152

retargeting (remarketing)campaigns, 145–147

Costco, 106Costs, of marketing. See Marketing

spendingCross-selling, 138–139Customer journey, 103–114

first value, 105–107“give to get”, 103, 155life-cycle marketing strategies,

109–114, 110new value, 108–109overview, 103–105, 104recommendations made during

customer life cycle, 140–141recurring value, 107–108

Customer personas, 93–101avoiding mistakes when using

clusters, 100insights, 101overview, 93–94types of clusters, 94–99, 96, 97, 98using clusters to improve

acquisition, 99–100Customer profiles, 43–61

data analysis preparation,50–54, 51

data collection design,45–47, 46

data collection types, 47, 47–57,48

overview, 43–45questions to use for data, 57–61

working with IT on dataintegration, 43, 54–57, 55

Customer relationship management(CRM) systems, 203–204

Customerscustomer-centricity, 224customer value path, 83–84, 84power of customer equity,

8–10, 11predictive marketing popularity

and, 3predictive marketing use cases,

11–12, 13privacy of, 215–220questions for data collection,

58–60See also Behavior of customers;

Conversion; Customerjourney; Customer personas;Customer profiles; Growth;Lifetime value; Predictiveanalytics; Reactivation;Retention; Value-basedmarketing

Data analysiscustomer control and, 218need for, 20, 20–21preparing data for, 50–54, 51vendor support for, 193–194

Data collectioncleansing and preparation, 37need for, 20, 20–21vendor support for, 191–193

Data Driven Marketing (Jeffrey),16–17

Data integrationcheckllist for, 189need for, 18–21, 20

Data management platforms(DMPs), 202

240 Index

Data Protection Directive (EuropeanUnion), 219

Data Security Standard, PaymentCredit Card Industry(PCI DSS), 219

Deduplication, 53–54Deighton, John, 120Delivery Point Validation (DPV), 52Design principles, for data collection,

45–47, 46Direct mail, for welcome

campaigns, 158Discount “junkies”, 24Discounts, likelihood to buy and,

126–128, 127Do-it-yourself predictive marketing,

197–198Dunn and Bradstreet, 234–235Dursun, Bulent, 6

Early adopterspredictive marketing popularity

and, 3predictive marketing value and,

16–17Earthlink, 17Einstein (First Union Bank), 122Email marketing

for abandoned cart campaigns,148–149

delivering, 194email service providers (ESPs),

202–203, 209frequency of email, 133, 139–136,

134, 136, 187likelihood to engage models,

130–136, 132, 133, 134, 136privacy issues and, 215–220retargeting with, 145–147validating addresses, 53

Engagement propensity model,222–223

Entenmann’s, 121Enthusiasts (email subscribers),

131–136, 132, 133, 134, 142

Facebook, look-alike targetingexample, 152, 158–153

Fallen from grace segment(high-value customers), 163

Feature generation/extraction, 38–39First Chicago Corporation, 122First party data, 45First-time buyers, likelihood to

buy, 124–125First Union Bank, 122First value, 105–107Fordham, Mark, 175–176Forrester, 142Free membership offers, 23Freemium business model, 106, 142Frequency, of email, 133, 139–136,

134, 136, 187Fuzzy matching, 53–54

GameStop, 165–167Gaming industry, 4–5Gaston, Katie, 175–176Gates, Bill, 16Gender, data collection about,

233–234Get Elastic, 217‘‘Give to get”, 103, 155Google Adwords, 47, 150Government legislation, privacy

issues and, 219–220Growth, 155–167

customer appreciation campaigns,161–163

initial transaction and, 155–156,156

loyalty programs, 163–166managing customers and, 70, 71negative churn concept and, 170

Index 241

new product introductions,160–161

omni-channel marketing, 166–167post purchase programs, 157–159replenishment campaigns and

repeat purchase programs,159–160

value-based marketing formedium-value customers,120–121

welcome programs, 157–158

Harrah’s Entertainment, 5Harte Hanks, 234–235Harvard Business Review, 120, 121High-value customers

customer appreciation campaignsfor, 161–163

marketing spending on, 86, 92–93,91

value-based marketing for,115–120, 116, 117, 118, 119

Historical lifetime value (LTV),64–65

Home addresses, data collectionabout, 231–232

Homejoy, 158Honda, 121Householding, 230–231

iBeacon, 53Identifiable information, of

customers, 216Imputation, 38Inactive customers, strategies for,

113–114Information technology (IT),

197–207advanced analytics, 204campaign management and

marketing cloud options,199–200

customer relationship management(CRM), 203–204

data integration and, 43,54–57, 55

data management platforms(DMPs), 202

do-it-yourself predictivemarketing, 197–198

email service providers (ESPs),202–203, 209

getting started with predictivemarketing, 205–207

goals for, 204–205machine learning, 7–8, 16, 27–28outsourcing to marketing service

providers (MSPs), 198–199overview, 201predictive marketing popularity

and, 3, 17–20technical capabilities for predictive

marketing, 187–191vendor support and, 194–196web analytics, 202

Initial transaction, importance of,155–156, 156

IP numbers, data collection about,231–232

Jeffrey, Mark, 16–18

Karabuk, Tulin, 94Kayak, 139Keyword-to-contact

recommendation model, 223K.I.D.S./Fashion Delivers, 106

Lapsed customersstrategies for, 107value-based marketing for,

118–119, 119See also Reactivation‘‘Last mile problem”, 40–41

242 Index

Last-touch attribution,89–92, 91

Lead scoring, predictive,128–130, 129, 130

Legal issues, of privacy, 219–220Lifetime value, 63–74

defined, 64historical lifetime value

(LTV), 64–65increasing, for all customers,

73–74, 74increasing, for one customer, 70,

70–73optimizing (See Marketing

spending)overview, 63–64predicted customer value, 66–68upside lifetime value, 68–70See also Customer journey;

Growth; RetentionLikelihood models

defined, 123–124likelihood to engage models,

130–136, 132, 133, 134, 136predictions, 124–130, 127, 129,

130See also Propensity models

LinkedIn, 113, 235Linking, 53–54Location, data collection about,

231–232Long-term email revenue, 133,

133–136, 134, 136Look-alike targeting

defined, 151–152Facebook example, 152, 152–153optimizing for similarity or reach,

153–154Low-value customers, value-based

marketing for, 115–119, 116,117, 118, 119, 122

Loyalty programscustomer profile data collection, 49growing customer value and,

163–166rise of predictive marketing and,

5–6Lululemon, 106

Machine learning, 7–8, 16, 27–28Mainstreet (email subscribers),

131–136, 132, 133, 134, 142‘‘Manage Marketing by the

Customer Equity Test”(Blattberg, Deighton), 120

Marketing, as art and science,211–212

Marketing funnelengineering, 81, 81–82growing customer value and,

155–156, 156by life cycle, 82, 82–83

Marketing service providers (MSPs),outsourcing to, 198–199

Marketing spending, 77–92for acquisition, retention,

reactivation, 78–86, 79, 80,81, 82, 83, 84, 85

channels for last touch attribution,89–92, 91

channels that bring high-valuecustomers, 88–89

conversion rate and, 81, 87–83,82, 83

for customer retention, 177–178differentiating, based on customer

value, 86, 86–87overview, 77–78products that bring high-value

customers, 87, 88to service low-value

customers, 122Marshall School of Business, 215

Index 243

Mavi, 5–6, 17, 49, 121, 164–165McDonald’s, 120‘‘Mean Stinks” campaign (Procter &

Gamble), 105Medium-value customers,

value-based marketing for,115–119, 116, 117, 118, 119,120–121

Meetings, data collection and,232–233

Metaphonic algorithms, 52Micro Warehouse, 198Mobile technology, customer

profiles and, 232Moosejaw, 111–112MSA/region append, 52Multitouch attribution, 89–92, 91

National Change of Address(NCOA), 52

Negative churn concept, 170Netflix, 4Newbies (email subscribers),

131–136, 132, 133, 134, 136New customers, strategies for,

111–112New product introductions, 160–161New value, 108–109New York Times, 125Next-sell recommendations, 140Nippon Telegraph and Telephone

(NTT), 211‘‘No free lunch” theorem, 39

Obama, Barack, 125Old school segment (high-value

customers), 163Omni-channel marketing, 166–167100% Pure, 167, 231Oner, Elif, 6Orange, 71–73Order of performance, 90

Organizational capabilities, checklistfor, 185–187

Outlier detection, 37–38Outsourcing

data science and, 206to marketing service providers

(MSPs), 198–199

PCI DSS (Data Security Standard,Payment Credit CardIndustry), 219

‘‘Personalization versus Privacy”(Marshall School ofBusiness), 215

Personalized recommendations.See Recommendations

PetCareRx, 114, 123Phantoms (email subscribers),

131–136, 132, 133, 134, 142Phased data integration strategy, 47,

47–48Pingree, Dan, 112Poolcycle management framework,

73, 74Post purchase programs, 157–159PowerUp Rewards (GameStop),

165–166Predicted customer value, 66–68Predictive analytics, 23–41

checklist for, 190defined, 24–25overview, 23–24process, 36, 36–41reinforcement learning and

collaborative filtering,33–36, 35

supervised learning, 28–32, 29, 30unsupervised learning, 25–28, 27

Predictive clustering model. SeeClustering

Predictive marketing, 3–21building blocks for, 20, 20–21

244 Index

Predictive marketing (continued)careers in, 209–213future of, 221–227power of customer equity,

8–10, 11privacy issues, 215–220relationship marketing with,

6–8, 8rise of, 3–6, 14–20use cases, 11–12, 13See also Behavior of customers;

Checklist of predictivemarketing capabilities;Conversion; Customerjourney; Customer personas;Customer profiles; Growth;Lifetime value; Predictiveanalytics; Reactivation;Retention; Value-basedmarketing

Pricing optimization model, 223Privacy, 215–220Proactive retention management,

175–180Procter & Gamble, 105Products

finding products that bringhigh-value customers, 87, 88

new product introductions,160–161

product-based clusters, 94,95, 96

product-to-productrecommendations, 34–35, 35,141–142

product-to-user typesrecommendations, 35

questions for data collection, 61Propensity models, 123–137

defined, 28–29, 123–124likelihood to buy predictions,

124–130, 127, 129, 136

likelihood to engage models,130–136, 132, 133,134, 136

propensity deciles, 29–31, 30RFM modeling compared

to, 31–32supervised learning, 28–32, 29, 30training and testing periods,

29, 29Prospective customers

converting, 145–151look-alike targeting, 151–154, 152strategies for, 109–111

Purchasesdata on, 47–50, 48making recommendations at time

of, 138–139

Reactivationcampaigns for, 180–182of inactive customers, 74marketing spending optimization

for, 78–86, 79, 80, 81, 82,83, 84, 85

value-based marketing for,118–119, 119

Rebecca Minkoff (stores), 221Recommendations, 137–144

channels for, 144choosing customers or segments,

138–141content and, 143–144overview, 33–34, 137–138types of models, 34–36, 35understanding customer

context, 141–143Recurring value, 107–108Reichheld, Frederick F., 121Reinforcement learning

collaborative filtering and,33–36, 35

defined, 25

Index 245

Repeat customerslikelihood to buy, 126strategies for, 112–113

Repeat purchase programs, 159–160Replenishment programs, 24,

159–160Response models. See Propensity

modelsRetargeting (remarketing) campaigns

overview, 145triggers for, 146–147

Retention, 169–182business models and, 171, 171churn differences, 172, 172–174,

173churn management programs,

174–175, 175customer management and, 70, 71data collection and, 49marketing spending optimization

for, 78–86, 79, 80, 81, 82,83, 84, 85

negative churn concept, 170proactive retention management,

175–180reactivation campaigns, 180–182understanding retention rate, 169value-based marketing for

high-value customers,119–120

Returnsby customers, 25data collection and, 233

Reviews, data collection and, 233Reward addict segment (high-value

customers), 162–163RFID, 221–222Rosling, Hans, 26

Safe Harbor Principles, 219Sainsbury Stores, 17

Sales, questions for datacollection, 57–58

Savings Catcher (Walmart), 50Search, abandoned, 149–150Secret (Procter & Gamble), 105Segmentation

recommendations andchoosing customersor segments, 138–141

vendor support for, 193–194See also Clustering; Customer

profilesShaklee, 21Shazam, 113–114, 186Shehata, George, 21Short-term email revenue, 133,

139–136, 134, 142Size data, collection of, 234–235Size of wallet, 68–70,

178, 223Sleepies (email subscribers),

131–136, 132, 133,134, 136

SmileTrain, 111Social interactions, data collection

and, 232–233Spotify, 142–143‘‘Sunk costs”, 85Supervised learning

defined, 25propensity models, 28–32

Surveys, data collection and, 233

Targetingimproving precision of, 11–12vendor support for, 193–194

Tchin Tchin campaign (AlainAfflelou), 139

TechCrunch, 142Third party data, 43, 235Total Addressable Market

(TAM), 223

246 Index

Total Rewards (Harrah’sEntertainment), 5

Transition matrix, 117,117–118, 118

Uncommon Goods, 131Unidentifiable information, of

customers, 216Unsupervised learning

clustering models, 25–28, 27defined, 25

Upselling, 138–139Upside lifetime value, 68–70U.S. census data, collection of, 234User-to-product recommendations,

142US-EU Safe Harbor Privacy Principles,

219U.S. Postal Service, 52

Validation, of customer data, 51–54Value-based marketing, 115–122

defined, 115growing medium-value customers,

120–121overview, 115–119, 116, 117,

118, 125

reducing costs to service low-valuecustomers, 122

retaining high-value customers,119–120

Value migration, 64–65, 174Variable contribution margin, 86Vendors, questions to ask of,

191–196Vertical data, collection of,

234–235VIP customers. See High-value

customers

Wallet analysis, 68–70, 178, 223Walmart, 50Warning signs, of customer

unhappiness, 176–180Web analytics, 202Web behavior, data collection about,

229–230Welcome programs, 157–158‘‘Whales”, 120

York, Jerry, 198

Zappos, 7, 146, 164Zendesk, 110

WILEY END USER LICENSE AGREEMENTGo to www.wiley.com/go/eula to access Wiley’s ebook EULA.

http://www.wiley.com/go/eula

Predictive Marketing

Documents

Transcript of Predictive Marketing